/* * Copyright (c) 2017-2021, Intel Corporation * * Permission is hereby granted, free of charge, to any person obtaining a * copy of this software and associated documentation files (the "Software"), * to deal in the Software without restriction, including without limitation * the rights to use, copy, modify, merge, publish, distribute, sublicense, * and/or sell copies of the Software, and to permit persons to whom the * Software is furnished to do so, subject to the following conditions: * * The above copyright notice and this permission notice shall be included * in all copies or substantial portions of the Software. * * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR * OTHER DEALINGS IN THE SOFTWARE. */ //! //! \file cm_hal.cpp //! \brief HAL Layer for CM Component //! #include "mos_os.h" #include "cm_hal.h" #include "media_interfaces_cmhal.h" #include "media_interfaces_mhw.h" #include "cm_common.h" #include "cm_hal_vebox.h" #include "cm_mem.h" #include "renderhal_platform_interface.h" #include "cm_execution_adv.h" #include "cm_extension_creator.h" #include "mos_interface.h" #define INDEX_ALIGN(index, elemperIndex, base) ((index * elemperIndex)/base + ( (index *elemperIndex % base))? 1:0) //---------------------------------- //| CM scoreboard XY //---------------------------------- struct CM_HAL_SCOREBOARD_XY { int32_t x; int32_t y; }; typedef CM_HAL_SCOREBOARD_XY *PCM_HAL_SCOREBOARD_XY; //--------------------------------------- //| CM scoreboard XY with mask //--------------------------------------- struct CM_HAL_SCOREBOARD_XY_MASK { int32_t x; int32_t y; uint8_t mask; uint8_t resetMask; }; typedef CM_HAL_SCOREBOARD_XY_MASK *PCM_HAL_SCOREBOARD_XY_MASK; //------------------------------------------------------------------------------ //| CM kernel slice and subslice being assigned to (for EnqueueWithHints) //------------------------------------------------------------------------------ struct CM_HAL_KERNEL_SLICE_SUBSLICE { uint32_t slice; uint32_t subSlice; }; typedef CM_HAL_KERNEL_SLICE_SUBSLICE *PCM_HAL_KERNEL_SLICE_SUBSLICE; //------------------------------------------------------------------------------ //| CM kernel information for EnqueueWithHints to assign subslice //------------------------------------------------------------------------------ struct CM_HAL_KERNEL_SUBSLICE_INFO { uint32_t numSubSlices; uint32_t counter; PCM_HAL_KERNEL_SLICE_SUBSLICE destination; }; typedef CM_HAL_KERNEL_SUBSLICE_INFO *PCM_HAL_KERNEL_SUBSLICE_INFO; // forward declaration int32_t HalCm_InsertCloneKernel( PCM_HAL_STATE state, PCM_HAL_KERNEL_PARAM kernelParam, PRENDERHAL_KRN_ALLOCATION &kernelAllocation); #if MDF_COMMAND_BUFFER_DUMP extern int32_t HalCm_InitDumpCommandBuffer(PCM_HAL_STATE state); extern int32_t HalCm_DumpCommadBuffer(PCM_HAL_STATE state, PMOS_COMMAND_BUFFER cmdBuffer, int offsetSurfaceState, size_t sizeOfSurfaceState); #endif #if MDF_CURBE_DATA_DUMP extern int32_t HalCm_InitDumpCurbeData(PCM_HAL_STATE state); extern int32_t HalCm_DumpCurbeData(PCM_HAL_STATE state); #endif #if MDF_SURFACE_CONTENT_DUMP extern int32_t HalCm_InitSurfaceDump(PCM_HAL_STATE state); #endif #if MDF_SURFACE_STATE_DUMP extern int32_t HalCm_InitDumpSurfaceState(PCM_HAL_STATE state); extern int32_t HalCm_DumpSurfaceState(PCM_HAL_STATE state, int offsetSurfaceState, size_t sizeOfSurfaceState); #endif #if MDF_INTERFACE_DESCRIPTOR_DATA_DUMP extern int32_t HalCm_InitDumpInterfaceDescriporData(PCM_HAL_STATE state); extern int32_t HalCm_DumpInterfaceDescriptorData(PCM_HAL_STATE state); #endif extern uint64_t HalCm_GetTsFrequency(PMOS_INTERFACE pOsInterface); //=========================================================== //*----------------------------------------------------------------------------- //| Purpose: Align to the next power of 2 //| Returns: Aligned data //| Reference: http://graphics.stanford.edu/~seander/bithacks.html#DetermineIfPowerOf2 //*----------------------------------------------------------------------------- __inline uint32_t HalCm_GetPow2Aligned(uint32_t d) { CM_ASSERT(d > 0); // subtract the number first --d; d |= d >> 1; d |= d >> 2; d |= d >> 4; d |= d >> 8; d |= d >> 16; return ++d; } //*----------------------------------------------------------------------------- //| Purpose: Checks if Task has any thread arguments //| Returns: True if task has any thread arguments, false otherwise //*----------------------------------------------------------------------------- bool HalCm_GetTaskHasThreadArg(PCM_HAL_KERNEL_PARAM *kernels, uint32_t numKernels) { PCM_HAL_KERNEL_PARAM kernelParam; PCM_HAL_KERNEL_ARG_PARAM argParam; bool threadArgExists = false; for( uint32_t krn = 0; krn < numKernels; krn++) { kernelParam = kernels[krn]; for(uint32_t argIndex = 0; argIndex < kernelParam->numArgs; argIndex++) { argParam = &kernelParam->argParams[argIndex]; if( argParam->perThread ) { threadArgExists = true; break; } } if( threadArgExists ) break; } return threadArgExists; } //*----------------------------------------------------------------------------- //| Purpose: Allocate Timestamp Resource //| Returns: Result of the operation //*----------------------------------------------------------------------------- MOS_STATUS HalCm_AllocateTsResource( PCM_HAL_STATE state) // [in] Pointer to CM HAL State { MOS_STATUS eStatus = MOS_STATUS_SUCCESS; uint32_t size; PMOS_INTERFACE osInterface; MOS_ALLOC_GFXRES_PARAMS allocParams; MOS_LOCK_PARAMS lockFlags; osInterface = state->osInterface; CM_CHK_NULL_GOTOFINISH_MOSERROR(osInterface); size = state->cmHalInterface->GetTimeStampResourceSize() * state->cmDeviceParam.maxTasks; // allocate render engine Ts Resource MOS_ZeroMemory(&allocParams, sizeof(MOS_ALLOC_GFXRES_PARAMS)); allocParams.Type = MOS_GFXRES_BUFFER; allocParams.dwBytes = size; allocParams.Format = Format_Buffer; //used in RenderHal_OsAllocateResource_Linux allocParams.TileType= MOS_TILE_LINEAR; allocParams.pBufName = "TsResource"; CM_CHK_HRESULT_GOTOFINISH_MOSERROR( osInterface->pfnAllocateResource(osInterface, &allocParams, &state->renderTimeStampResource.osResource)); // RegisterResource will be called in AddResourceToHWCmd. It is not allowed to be called by hal explicitly if (!osInterface->apoMosEnabled) { CM_CHK_MOSSTATUS_GOTOFINISH( osInterface->pfnRegisterResource(osInterface, &state->renderTimeStampResource.osResource, true, true)); } osInterface->pfnSkipResourceSync(&state->renderTimeStampResource.osResource); // Lock the Resource MOS_ZeroMemory(&lockFlags, sizeof(MOS_LOCK_PARAMS)); lockFlags.ReadOnly = 1; lockFlags.ForceCached = true; state->renderTimeStampResource.data = (uint8_t*)osInterface->pfnLockResource( osInterface, &state->renderTimeStampResource.osResource, &lockFlags); CM_CHK_NULL_GOTOFINISH_MOSERROR(state->renderTimeStampResource.data); state->renderTimeStampResource.locked = true; //allocated for vebox TS resource MOS_ZeroMemory(&allocParams, sizeof(MOS_ALLOC_GFXRES_PARAMS)); allocParams.Type = MOS_GFXRES_BUFFER; allocParams.dwBytes = size; allocParams.Format = Format_Buffer; //used in RenderHal_OsAllocateResource_Linux allocParams.TileType = MOS_TILE_LINEAR; allocParams.pBufName = "TsResource"; CM_CHK_HRESULT_GOTOFINISH_MOSERROR(osInterface->pfnAllocateResource( osInterface, &allocParams, &state->veboxTimeStampResource.osResource)); // Lock the Resource MOS_ZeroMemory(&lockFlags, sizeof(MOS_LOCK_PARAMS)); lockFlags.ReadOnly = 1; lockFlags.ForceCached = true; state->veboxTimeStampResource.data = (uint8_t*)osInterface->pfnLockResource( osInterface, &state->veboxTimeStampResource.osResource, &lockFlags); CM_CHK_NULL_GOTOFINISH_MOSERROR(state->veboxTimeStampResource.data); state->veboxTimeStampResource.locked = true; finish: return eStatus; } //! \brief Allocate tracker resource //! \param [in] state //! Pointer to CM_HAL_STATE structure //! \return MOS_STATUS MOS_STATUS HalCm_AllocateTrackerResource( PCM_HAL_STATE state) { MOS_STATUS eStatus = MOS_STATUS_SUCCESS; MOS_ALLOC_GFXRES_PARAMS allocParamsLinearBuffer; MOS_LOCK_PARAMS lockFlags; PMOS_INTERFACE osInterface; PRENDERHAL_INTERFACE renderHal; osInterface = state->osInterface; renderHal = state->renderHal; CM_CHK_NULL_GOTOFINISH_MOSERROR(osInterface); // Tracker producer for RENDER engine renderHal->trackerProducer.Initialize(osInterface); // Tracker resource for VeBox engine osInterface->pfnResetResource(&renderHal->veBoxTrackerRes.osResource); MOS_ZeroMemory(&allocParamsLinearBuffer, sizeof(MOS_ALLOC_GFXRES_PARAMS)); allocParamsLinearBuffer.Type = MOS_GFXRES_BUFFER; allocParamsLinearBuffer.TileType = MOS_TILE_LINEAR; allocParamsLinearBuffer.Format = Format_Buffer; allocParamsLinearBuffer.dwBytes = MHW_CACHELINE_SIZE; allocParamsLinearBuffer.pBufName = "VeboxTrackerRes"; CM_CHK_HRESULT_GOTOFINISH_MOSERROR(osInterface->pfnAllocateResource( osInterface, &allocParamsLinearBuffer, &renderHal->veBoxTrackerRes.osResource)); // Lock the Resource MOS_ZeroMemory(&lockFlags, sizeof(MOS_LOCK_PARAMS)); lockFlags.ReadOnly = 1; lockFlags.ForceCached = true; renderHal->veBoxTrackerRes.data = (uint32_t*)osInterface->pfnLockResource( osInterface, &renderHal->veBoxTrackerRes.osResource, &lockFlags); CM_CHK_NULL_GOTOFINISH_MOSERROR(renderHal->veBoxTrackerRes.data); *(renderHal->veBoxTrackerRes.data) = MemoryBlock::m_invalidTrackerId; renderHal->veBoxTrackerRes.currentTrackerId = 1; renderHal->veBoxTrackerRes.locked = true; finish: return eStatus; } //! \brief Initialize dynamic state heap //! \param [in] state //! Pointer to CM_HAL_STATE structure //! \param [in] heapParam //! Pointer to CM_HAL_HEAP_PARAM structure //! \return MOS_STATUS MOS_STATUS HalCm_InitializeDynamicStateHeaps( PCM_HAL_STATE state, CM_HAL_HEAP_PARAM *heapParam) { MOS_STATUS eStatus = MOS_STATUS_SUCCESS; HeapManager* dgsHeap = state->renderHal->dgsheapManager; CM_CHK_NULL_GOTOFINISH_MOSERROR(heapParam); dgsHeap = MOS_New(HeapManager); CM_CHK_NULL_GOTOFINISH_MOSERROR(dgsHeap); CM_CHK_MOSSTATUS_GOTOFINISH(dgsHeap->RegisterOsInterface(state->osInterface)); dgsHeap->SetDefaultBehavior(heapParam->behaviorGSH); CM_CHK_MOSSTATUS_GOTOFINISH(dgsHeap->SetInitialHeapSize(heapParam->initialSizeGSH)); CM_CHK_MOSSTATUS_GOTOFINISH(dgsHeap->SetExtendHeapSize(heapParam->extendSizeGSH)); CM_CHK_MOSSTATUS_GOTOFINISH(dgsHeap->RegisterTrackerProducer(heapParam->trackerProducer)); // lock the heap in the beginning, so cpu doesn't need to wait gpu finishing occupying it to lock it again CM_CHK_MOSSTATUS_GOTOFINISH(dgsHeap->LockHeapsOnAllocate()); state->renderHal->dgsheapManager = dgsHeap; finish: return eStatus; } //*----------------------------------------------------------------------------- //| Purpose: Free Timestamp Resource //| Returns: Result of the operation //*----------------------------------------------------------------------------- __inline void HalCm_FreeTsResource( PCM_HAL_STATE state) // [in] Pointer to CM HAL State { PMOS_INTERFACE osInterface; MOS_STATUS hr; MOS_GFXRES_FREE_FLAGS resFreeFlags = {0}; resFreeFlags.AssumeNotInUse = 1; osInterface = state->osInterface; if (!Mos_ResourceIsNull(&state->renderTimeStampResource.osResource)) { if (state->renderTimeStampResource.locked) { hr = (MOS_STATUS)osInterface->pfnUnlockResource( osInterface, &state->renderTimeStampResource.osResource); CM_ASSERT(hr == MOS_STATUS_SUCCESS); } osInterface->pfnFreeResourceWithFlag( osInterface, &state->renderTimeStampResource.osResource, resFreeFlags.Value); } //free vebox TS resource if (!Mos_ResourceIsNull(&state->veboxTimeStampResource.osResource)) { if (state->veboxTimeStampResource.locked) { hr = (MOS_STATUS)osInterface->pfnUnlockResource( osInterface, &state->veboxTimeStampResource.osResource); CM_ASSERT(hr == MOS_STATUS_SUCCESS); } osInterface->pfnFreeResourceWithFlag( osInterface, &state->veboxTimeStampResource.osResource, resFreeFlags.Value); } } //! \brief Free tracker resource //! \param PCM_HAL_STATE state //! [in] Pointer to CM_HAL_STATE structure //! \return void __inline void HalCm_FreeTrackerResources( PCM_HAL_STATE state) // [in] Pointer to CM HAL State { PMOS_INTERFACE osInterface; MOS_STATUS hr; MOS_GFXRES_FREE_FLAGS resFreeFlags = {0}; resFreeFlags.AssumeNotInUse = 1; osInterface = state->osInterface; if (!Mos_ResourceIsNull(&state->renderHal->veBoxTrackerRes.osResource)) { if (state->renderHal->veBoxTrackerRes.locked) { hr = (MOS_STATUS)osInterface->pfnUnlockResource( osInterface, &state->renderHal->veBoxTrackerRes.osResource); CM_ASSERT(hr == MOS_STATUS_SUCCESS); } osInterface->pfnFreeResourceWithFlag( osInterface, &state->renderHal->veBoxTrackerRes.osResource, resFreeFlags.Value); } } //*----------------------------------------------------------------------------- //| Purpose: Allocate CSR Resource //| Returns: Result of the operation //*----------------------------------------------------------------------------- MOS_STATUS HalCm_AllocateCSRResource( PCM_HAL_STATE state) // [in] Pointer to CM HAL State { PMOS_INTERFACE osInterface = state->osInterface; MOS_STATUS eStatus = MOS_STATUS_SUCCESS; uint32_t size; MOS_ALLOC_GFXRES_PARAMS allocParams; //Enable Mid-thread state->renderHal->pfnEnableGpgpuMiddleThreadPreemption(state->renderHal); size = CM_CSR_SURFACE_SIZE; MOS_ZeroMemory(&allocParams, sizeof(MOS_ALLOC_GFXRES_PARAMS)); allocParams.Type = MOS_GFXRES_BUFFER; allocParams.dwBytes = size; allocParams.Format = Format_RAW; //used in VpHal_OsAllocateResource_Linux allocParams.TileType = MOS_TILE_LINEAR; allocParams.pBufName = "CSRResource"; CM_CHK_HRESULT_GOTOFINISH_MOSERROR(osInterface->pfnAllocateResource( osInterface, &allocParams, &state->csrResource)); osInterface->pfnSkipResourceSync(&state->csrResource); finish: return eStatus; } //*----------------------------------------------------------------------------- //| Purpose: Allocate Sip Resource //| Returns: Result of the operation //*----------------------------------------------------------------------------- MOS_STATUS HalCm_AllocateSipResource( PCM_HAL_STATE state) // [in] Pointer to CM HAL State { PMOS_INTERFACE osInterface = state->osInterface; MOS_STATUS eStatus = MOS_STATUS_SUCCESS; uint32_t size; MOS_ALLOC_GFXRES_PARAMS allocParams; MOS_LOCK_PARAMS lockFlags; size = CM_DEBUG_SURFACE_SIZE; MOS_ZeroMemory(&allocParams, sizeof(MOS_ALLOC_GFXRES_PARAMS)); allocParams.Type = MOS_GFXRES_BUFFER; allocParams.dwBytes = size; allocParams.Format = Format_Buffer; //used in RenderHal_OsAllocateResource_Linux allocParams.TileType = MOS_TILE_LINEAR; allocParams.pBufName = "SipResource"; CM_CHK_HRESULT_GOTOFINISH_MOSERROR(osInterface->pfnAllocateResource( osInterface, &allocParams, &state->sipResource.osResource)); // Lock the Resource MOS_ZeroMemory(&lockFlags, sizeof(MOS_LOCK_PARAMS)); lockFlags.ReadOnly = 1; lockFlags.ForceCached = true; state->sipResource.data = (uint8_t*)osInterface->pfnLockResource( osInterface, &state->sipResource.osResource, &lockFlags); CM_CHK_NULL_GOTOFINISH_MOSERROR(state->sipResource.data); state->sipResource.locked = true; finish: return eStatus; } //*----------------------------------------------------------------------------- //| Purpose: Free CSR Resource //| Returns: Result of the operation //*----------------------------------------------------------------------------- __inline void HalCm_FreeCsrResource( PCM_HAL_STATE state) // [in] Pointer to CM HAL State { PMOS_INTERFACE osInterface = state->osInterface; MOS_GFXRES_FREE_FLAGS resFreeFlags = {0}; resFreeFlags.AssumeNotInUse = 1; if (!Mos_ResourceIsNull(&state->csrResource)) { osInterface->pfnFreeResourceWithFlag( osInterface, &state->csrResource, resFreeFlags.Value); } } //*----------------------------------------------------------------------------- //| Purpose: Free Sip Resource //| Returns: Result of the operation //*----------------------------------------------------------------------------- __inline void HalCm_FreeSipResource( PCM_HAL_STATE state) // [in] Pointer to CM HAL State { PMOS_INTERFACE osInterface = state->osInterface; MOS_STATUS hr = MOS_STATUS_SUCCESS; MOS_GFXRES_FREE_FLAGS resFreeFlags = {0}; resFreeFlags.AssumeNotInUse = 1; if (!Mos_ResourceIsNull(&state->sipResource.osResource)) { if (state->sipResource.locked) { hr = (MOS_STATUS)osInterface->pfnUnlockResource( osInterface, &state->sipResource.osResource); CM_ASSERT(hr == MOS_STATUS_SUCCESS); } osInterface->pfnFreeResourceWithFlag( osInterface, &state->sipResource.osResource, resFreeFlags.Value); } } //*----------------------------------------------------------------------------- //| Purpose: Sets Arg data in the buffer //| Returns: Result of the operation //*----------------------------------------------------------------------------- __inline void HalCm_SetArgData( PCM_HAL_KERNEL_ARG_PARAM argParam, uint32_t threadIndex, uint8_t *buffer) { uint8_t *dst; uint8_t *src; dst = buffer + argParam->payloadOffset; src = argParam->firstValue + (threadIndex * argParam->unitSize); MOS_SecureMemcpy(dst, argParam->unitSize, src, argParam->unitSize); } //*----------------------------------------------------------------------------- //| Purpose: Get the Buffer Entry //| Returns: Result of the operation. //*----------------------------------------------------------------------------- __inline MOS_STATUS HalCm_GetResourceUPEntry( PCM_HAL_STATE state, // [in] Pointer to CM State uint32_t handle, // [in] Handle PCM_HAL_SURFACE2D_UP_ENTRY *entryOut) // [out] Buffer Entry { MOS_STATUS eStatus; PCM_HAL_SURFACE2D_UP_ENTRY entry; eStatus = MOS_STATUS_SUCCESS; if (handle >= state->cmDeviceParam.max2DSurfaceUPTableSize) { eStatus = MOS_STATUS_INVALID_HANDLE; CM_ASSERTMESSAGE("Invalid handle '%d'", handle); goto finish; } entry = &state->surf2DUPTable[handle]; if (entry->width == 0) { eStatus = MOS_STATUS_INVALID_HANDLE; CM_ASSERTMESSAGE("handle '%d' is not set", handle); goto finish; } *entryOut = entry; finish: return eStatus; } //*----------------------------------------------------------------------------- //| Purpose: Get the Buffer Entry //| Returns: Result of the operation. //*----------------------------------------------------------------------------- __inline MOS_STATUS HalCm_GetBufferEntry( PCM_HAL_STATE state, // [in] Pointer to CM State uint32_t handle, // [in] Handle PCM_HAL_BUFFER_ENTRY *entryOut) // [out] Buffer Entry { MOS_STATUS eStatus; PCM_HAL_BUFFER_ENTRY entry; eStatus = MOS_STATUS_SUCCESS; if (handle >= state->cmDeviceParam.maxBufferTableSize) { eStatus = MOS_STATUS_INVALID_HANDLE; CM_ASSERTMESSAGE("Invalid handle '%d'", handle); goto finish; } entry = &state->bufferTable[handle]; if (entry->size == 0) { eStatus = MOS_STATUS_INVALID_HANDLE; CM_ASSERTMESSAGE("handle '%d' is not set", handle); goto finish; } *entryOut = entry; finish: return eStatus; } //*----------------------------------------------------------------------------- //| Purpose: Get the Surface2D Entry //| Returns: Result of the operation. //*----------------------------------------------------------------------------- __inline MOS_STATUS HalCm_GetSurface2DEntry( PCM_HAL_STATE state, // [in] Pointer to CM State uint32_t handle, // [in] Handle PCM_HAL_SURFACE2D_ENTRY *entryOut) // [out] Buffer Entry { MOS_STATUS eStatus; PCM_HAL_SURFACE2D_ENTRY entry; eStatus = MOS_STATUS_SUCCESS; if (handle >= state->cmDeviceParam.max2DSurfaceTableSize) { eStatus = MOS_STATUS_INVALID_HANDLE; CM_ASSERTMESSAGE("Invalid handle '%d'", handle); goto finish; } entry = &state->umdSurf2DTable[handle]; if ((entry->width == 0)||(entry->height == 0)) { eStatus = MOS_STATUS_INVALID_HANDLE; CM_ASSERTMESSAGE("handle '%d' is not set", handle); goto finish; } *entryOut = entry; finish: return eStatus; } //*----------------------------------------------------------------------------- //| Purpose: Get the 3D Entry //| Returns: Result of the operation. //*----------------------------------------------------------------------------- __inline MOS_STATUS HalCm_Get3DResourceEntry( PCM_HAL_STATE state, // [in] Pointer to CM State uint32_t handle, // [in] Handle PCM_HAL_3DRESOURCE_ENTRY *entryOut) // [out] Buffer Entry { MOS_STATUS eStatus; PCM_HAL_3DRESOURCE_ENTRY entry; eStatus = MOS_STATUS_SUCCESS; if (handle >= state->cmDeviceParam.max3DSurfaceTableSize) { eStatus = MOS_STATUS_INVALID_HANDLE; CM_ASSERTMESSAGE("Invalid handle '%d'", handle); goto finish; } entry = &state->surf3DTable[handle]; if (Mos_ResourceIsNull(&entry->osResource)) { eStatus = MOS_STATUS_INVALID_HANDLE; CM_ASSERTMESSAGE("3D handle '%d' is not set", handle); goto finish; } *entryOut = entry; finish: return eStatus; } //*----------------------------------------------------------------------------- //| Purpose: Allocates and sets up Task Param memory structure //| Return: true if enabled //| Note: A single layer of memory is allocated to avoid fragmentation //*----------------------------------------------------------------------------- MOS_STATUS HalCm_AllocateTables( PCM_HAL_STATE state) // [in] Pointer to HAL CM state { MOS_STATUS eStatus = MOS_STATUS_SUCCESS; PCM_HAL_DEVICE_PARAM deviceParam; uint8_t *pb; deviceParam = &state->cmDeviceParam; uint32_t lookUpTableSize = deviceParam->max2DSurfaceTableSize * sizeof(CMLOOKUP_ENTRY); uint32_t i2DSURFTableSize = deviceParam->max2DSurfaceTableSize * sizeof(CM_HAL_SURFACE2D_ENTRY); uint32_t bufferTableSize = deviceParam->maxBufferTableSize * sizeof(CM_HAL_BUFFER_ENTRY); uint32_t i2DSURFUPTableSize = deviceParam->max2DSurfaceUPTableSize * sizeof(CM_HAL_SURFACE2D_UP_ENTRY); uint32_t i3DSurfTableSize = deviceParam->max3DSurfaceTableSize * sizeof(CM_HAL_3DRESOURCE_ENTRY); uint32_t samplerTableSize = deviceParam->maxSamplerTableSize * sizeof(MHW_SAMPLER_STATE_PARAM); uint32_t sampler8x8TableSize = deviceParam->maxSampler8x8TableSize * sizeof(CM_HAL_SAMPLER_8X8_ENTRY); uint32_t taskStatusTableSize = deviceParam->maxTasks * sizeof(char); uint32_t bt2DIndexTableSize = deviceParam->max2DSurfaceTableSize * sizeof(CM_HAL_MULTI_USE_BTI_ENTRY); uint32_t bt2DUPIndexTableSize = deviceParam->max2DSurfaceUPTableSize * sizeof(CM_HAL_MULTI_USE_BTI_ENTRY); uint32_t bt3DIndexTableSize = deviceParam->max3DSurfaceTableSize * sizeof(CM_HAL_MULTI_USE_BTI_ENTRY); uint32_t btbufferIndexTableSize = deviceParam->maxBufferTableSize * sizeof(CM_HAL_MULTI_USE_BTI_ENTRY); uint32_t samplerIndexTableSize = deviceParam->maxSamplerTableSize * sizeof(char); uint32_t sampler8x8IndexTableSize = deviceParam->maxSampler8x8TableSize * sizeof(char); uint32_t size = lookUpTableSize + i2DSURFTableSize + bufferTableSize + i2DSURFUPTableSize + i3DSurfTableSize + samplerTableSize + sampler8x8TableSize + taskStatusTableSize + bt2DIndexTableSize + bt2DUPIndexTableSize + bt3DIndexTableSize + btbufferIndexTableSize + samplerIndexTableSize + sampler8x8IndexTableSize; state->tableMemories = MOS_AllocAndZeroMemory(size); CM_CHK_NULL_GOTOFINISH_MOSERROR(state->tableMemories); pb = (uint8_t*)state->tableMemories; state->surf2DTable = (PCMLOOKUP_ENTRY)pb; pb += lookUpTableSize; state->umdSurf2DTable = (PCM_HAL_SURFACE2D_ENTRY)pb; pb += i2DSURFTableSize; state->bufferTable = (PCM_HAL_BUFFER_ENTRY)pb; pb += bufferTableSize; state->surf2DUPTable = (PCM_HAL_SURFACE2D_UP_ENTRY)pb; pb += i2DSURFUPTableSize; state->surf3DTable = (PCM_HAL_3DRESOURCE_ENTRY)pb; pb += i3DSurfTableSize; state->samplerTable = (PMHW_SAMPLER_STATE_PARAM)pb; pb += samplerTableSize; state->sampler8x8Table = (PCM_HAL_SAMPLER_8X8_ENTRY)pb; pb += sampler8x8TableSize; state->taskStatusTable = (char *)pb; pb += taskStatusTableSize; state->bti2DIndexTable = (PCM_HAL_MULTI_USE_BTI_ENTRY)pb; pb += bt2DIndexTableSize; state->bti2DUPIndexTable = (PCM_HAL_MULTI_USE_BTI_ENTRY)pb; pb += bt2DUPIndexTableSize; state->bti3DIndexTable = (PCM_HAL_MULTI_USE_BTI_ENTRY)pb; pb += bt3DIndexTableSize; state->btiBufferIndexTable = (PCM_HAL_MULTI_USE_BTI_ENTRY)pb; pb += btbufferIndexTableSize; state->samplerIndexTable = (char *)pb; pb += samplerIndexTableSize; state->sampler8x8IndexTable = (char *)pb; pb += sampler8x8IndexTableSize; finish: return MOS_STATUS_SUCCESS; } //*----------------------------------------------------------------------------- //| Purpose: Adds a tag to distinguish between same kernel ID //| Used for batch buffer re-use when splitting large task into //| smaller pieces for EnqueueWithHints //| Using bits [48:42] from kernel ID for extra tag //| Returns: Result of the operation //*----------------------------------------------------------------------------- MOS_STATUS HalCm_AddKernelIDTag( PCM_HAL_KERNEL_PARAM *pKernels, uint32_t numKernels, uint32_t numTasks, uint32_t numCurrentTask) { uint32_t i; uint64_t tmpNumTasks; uint64_t tmpNumCurrentTask; uint64_t tmpNumTasksMask; uint64_t tmpNumCurrentTaskMask; tmpNumTasks = numTasks; tmpNumCurrentTask = numCurrentTask; tmpNumTasksMask = tmpNumTasks << 45; tmpNumCurrentTaskMask = tmpNumCurrentTask << 42; for( i = 0; i < numKernels; ++i ) { pKernels[i]->kernelId |= tmpNumTasksMask; pKernels[i]->kernelId |= tmpNumCurrentTaskMask; } return MOS_STATUS_SUCCESS; } //*----------------------------------------------------------------------------- //| Purpose: Gets the Batch Buffer for rendering. If needed, de-allocate / //| allocate the memory for BB //| Returns: Result of the operation //*----------------------------------------------------------------------------- MOS_STATUS HalCm_GetBatchBuffer( PCM_HAL_STATE state, // [in] Pointer to CM State uint32_t numKernels, // [in] Number of Kernels PCM_HAL_KERNEL_PARAM *kernels, // [in] Array for kernel data PMHW_BATCH_BUFFER *batchBufferOut) // [out] Batch Buffer Out { MOS_STATUS eStatus; PMHW_BATCH_BUFFER batchBuffer = nullptr; PRENDERHAL_INTERFACE renderHal; int32_t size; uint32_t i; uint32_t j; uint32_t k; int32_t freeIdx; uint64_t kernelIds[CM_MAX_KERNELS_PER_TASK]; uint64_t kernelParamsIds[CM_MAX_KERNELS_PER_TASK]; CM_HAL_BB_DIRTY_STATUS bbDirtyStatus; PCM_HAL_BB_ARGS bbcmArgs; eStatus = MOS_STATUS_SUCCESS; renderHal = state->renderHal; freeIdx = CM_INVALID_INDEX; bbDirtyStatus = CM_HAL_BB_CLEAN; // Align the Batch Buffer size to power of 2 size = HalCm_GetPow2Aligned(state->taskParam->batchBufferSize); MOS_ZeroMemory(&kernelIds, CM_MAX_KERNELS_PER_TASK * sizeof(uint64_t)); MOS_ZeroMemory(&kernelParamsIds, CM_MAX_KERNELS_PER_TASK * sizeof(uint64_t)); //Sanity check for batch buffer if (size > CM_MAX_BB_SIZE) { eStatus = MOS_STATUS_EXCEED_MAX_BB_SIZE; CM_ASSERTMESSAGE("Batch Buffer Size exeeceds Max '%d'", size); goto finish; } for( i = 0; i < numKernels; ++i ) { // remove upper 16 bits used for kernel binary re-use in GSH kernelParamsIds[i] = ((kernels[i])->kernelId << 16 ) >> 16; } #if CM_BATCH_BUFFER_REUSE_ENABLE bbDirtyStatus = CM_HAL_BB_CLEAN; for (k = 0; k < numKernels; ++k) { if (kernels[k]->kernelThreadSpaceParam.bbDirtyStatus == CM_HAL_BB_DIRTY) { bbDirtyStatus = CM_HAL_BB_DIRTY; break; } } for (i = 0; i < (uint32_t)state->numBatchBuffers; i++) { batchBuffer = &state->batchBuffers[i]; CM_CHK_NULL_GOTOFINISH_MOSERROR(batchBuffer); CM_CHK_NULL_GOTOFINISH_MOSERROR(batchBuffer->pPrivateData); //if (!Mos_ResourceIsNull(&batchBuffer->OsResource) && (!batchBuffer->bBusy)) if (!Mos_ResourceIsNull(&batchBuffer->OsResource)) { MOS_FillMemory(kernelIds, sizeof(uint64_t)*CM_MAX_KERNELS_PER_TASK, 0); for (j = 0; j < numKernels; j ++) { kernelIds[j] = kernelParamsIds[j]; } bbcmArgs = (PCM_HAL_BB_ARGS)batchBuffer->pPrivateData; if (RtlEqualMemory(kernelIds, bbcmArgs->kernelIds, sizeof(uint64_t)*CM_MAX_KERNELS_PER_TASK)) { if( batchBuffer->bBusy && bbDirtyStatus == CM_HAL_BB_DIRTY ) { bbcmArgs->latest = false; } else if( bbcmArgs->latest == true ) { break; } } } } if (i < (uint32_t)state->numBatchBuffers) { CM_CHK_NULL_GOTOFINISH_MOSERROR(batchBuffer); CM_CHK_NULL_GOTOFINISH_MOSERROR(batchBuffer->pPrivateData); bbcmArgs = (PCM_HAL_BB_ARGS)batchBuffer->pPrivateData; bbcmArgs->refCount ++; batchBuffer->iCurrent = 0; batchBuffer->dwSyncTag = 0; batchBuffer->iRemaining = batchBuffer->iSize; *batchBufferOut = batchBuffer; eStatus = MOS_STATUS_SUCCESS; goto finish; } #endif for (i = 0; i < (uint32_t)state->numBatchBuffers; i++) { batchBuffer = &state->batchBuffers[i]; CM_CHK_NULL_GOTOFINISH_MOSERROR(batchBuffer); // No holes in the array of batch buffers if (Mos_ResourceIsNull(&batchBuffer->OsResource)) { freeIdx = i; break; } } if (freeIdx == CM_INVALID_INDEX) { for (i = 0; i < (uint32_t)state->numBatchBuffers; i++) { batchBuffer = &state->batchBuffers[i]; CM_CHK_NULL_GOTOFINISH_MOSERROR(batchBuffer); CM_CHK_NULL_GOTOFINISH_MOSERROR(batchBuffer->pPrivateData); bbcmArgs = (PCM_HAL_BB_ARGS)batchBuffer->pPrivateData; if (!batchBuffer->bBusy) { if (batchBuffer->iSize >= size) { batchBuffer->iCurrent = 0; batchBuffer->iRemaining = batchBuffer->iSize; batchBuffer->dwSyncTag = 0; bbcmArgs->refCount = 1; for (i = 0; i kernelIds[i] = kernelParamsIds[i]; } bbcmArgs->latest = true; *batchBufferOut = batchBuffer; eStatus = MOS_STATUS_SUCCESS; goto finish; } if (freeIdx == CM_INVALID_INDEX) { freeIdx = i; } } } } if (freeIdx == CM_INVALID_INDEX) { eStatus = MOS_STATUS_INVALID_PARAMETER; CM_ASSERTMESSAGE("No batch buffer available"); goto finish; } batchBuffer = &state->batchBuffers[freeIdx]; CM_CHK_NULL_GOTOFINISH_MOSERROR(batchBuffer); CM_CHK_NULL_GOTOFINISH_MOSERROR(batchBuffer->pPrivateData); bbcmArgs = (PCM_HAL_BB_ARGS)batchBuffer->pPrivateData; bbcmArgs->refCount = 1; for (i = 0; i kernelIds[i] = kernelParamsIds[i]; } bbcmArgs->latest = true; if (!Mos_ResourceIsNull(&batchBuffer->OsResource)) { // Deallocate Batch Buffer CM_CHK_MOSSTATUS_GOTOFINISH(renderHal->pfnFreeBB(renderHal, batchBuffer)); } // Allocate Batch Buffer CM_CHK_MOSSTATUS_GOTOFINISH(renderHal->pfnAllocateBB(renderHal, batchBuffer, size)); *batchBufferOut = batchBuffer; finish: return eStatus; } //*----------------------------------------------------------------------------- //| Purpose: Parse the Kernel and populate the Task Param structure //| Return: Result of the operation //*----------------------------------------------------------------------------- MOS_STATUS HalCm_ParseTask( PCM_HAL_STATE state, // [in] Pointer to HAL CM state PCM_HAL_EXEC_TASK_PARAM execParam) // [in] Pointer to Exec Task Param { MOS_STATUS eStatus; PCM_HAL_TASK_PARAM taskParam; PCM_HAL_KERNEL_PARAM kernelParam; uint32_t hdrSize; uint32_t totalThreads; uint32_t krn; uint32_t curbeOffset; PMHW_VFE_SCOREBOARD scoreboardParams; uint32_t hasThreadArg; bool nonstallingScoreboardEnable; CM_HAL_DEPENDENCY vfeDependencyInfo; PCM_HAL_KERNEL_THREADSPACE_PARAM kernelTSParam; uint32_t i, j, k; uint8_t reuseBBUpdateMask; bool bitIsSet; PCM_HAL_MASK_AND_RESET dependencyMask; uint32_t uSurfaceNumber; uint32_t uSurfaceIndex; bool threadArgExists; eStatus = MOS_STATUS_SUCCESS; curbeOffset = 0; totalThreads = 0; taskParam = state->taskParam; taskParam->batchBufferSize = 0; hasThreadArg = 0; nonstallingScoreboardEnable = true; reuseBBUpdateMask = 0; bitIsSet = false; threadArgExists = false; hdrSize = state->renderHal->pHwSizes->dwSizeMediaObjectHeaderCmd; taskParam->dependencyPattern = execParam->dependencyPattern; taskParam->threadSpaceWidth = execParam->threadSpaceWidth; taskParam->threadSpaceHeight = execParam->threadSpaceHeight; taskParam->walkingPattern = execParam->walkingPattern; taskParam->walkingParamsValid = execParam->walkingParamsValid; taskParam->dependencyVectorsValid = execParam->dependencyVectorsValid; if( taskParam->walkingParamsValid ) { taskParam->walkingParams = execParam->walkingParams; } if( taskParam->dependencyVectorsValid ) { taskParam->dependencyVectors = execParam->dependencyVectors; } taskParam->kernelDebugEnabled = (uint32_t)execParam->kernelDebugEnabled; //GT-PIN taskParam->surfEntryInfoArrays = execParam->surfEntryInfoArrays; taskParam->surfacePerBT = 0; taskParam->colorCountMinusOne = execParam->colorCountMinusOne; taskParam->mediaWalkerGroupSelect = execParam->mediaWalkerGroupSelect; if (execParam->threadCoordinates) { taskParam->threadCoordinates = execParam->threadCoordinates; } taskParam->dependencyMasks = execParam->dependencyMasks; taskParam->syncBitmap = execParam->syncBitmap; taskParam->conditionalEndBitmap = execParam->conditionalEndBitmap; MOS_SecureMemcpy(taskParam->conditionalEndInfo, sizeof(taskParam->conditionalEndInfo), execParam->conditionalEndInfo, sizeof(execParam->conditionalEndInfo)); taskParam->numKernels = execParam->numKernels; taskParam->taskConfig = execParam->taskConfig; state->walkerParams.CmWalkerEnable = true; state->renderHal->IsMDFLoad = (taskParam->taskConfig.turboBoostFlag == CM_TURBO_BOOST_ENABLE); for (krn = 0; krn < execParam->numKernels; krn++) { if ((execParam->kernels[krn] == nullptr) || (execParam->kernelSizes[krn] == 0)) { eStatus = MOS_STATUS_INVALID_PARAMETER; CM_ASSERTMESSAGE("Invalid Kernel data"); goto finish; } kernelParam = (PCM_HAL_KERNEL_PARAM)execParam->kernels[krn]; PCM_INDIRECT_SURFACE_INFO indirectSurfaceInfo = kernelParam->indirectDataParam.surfaceInfo; uSurfaceNumber = 0; if (kernelParam->indirectDataParam.surfaceCount) { uSurfaceIndex = 0; for (i = 0; i < kernelParam->indirectDataParam.surfaceCount; i++) { uSurfaceIndex = (indirectSurfaceInfo + i)->bindingTableIndex > uSurfaceIndex ? ((indirectSurfaceInfo + i)->bindingTableIndex + (indirectSurfaceInfo + i)->numBTIPerSurf - 1) : uSurfaceIndex; uSurfaceNumber = uSurfaceNumber + (indirectSurfaceInfo + i)->numBTIPerSurf; } taskParam->surfacePerBT = taskParam->surfacePerBT > uSurfaceIndex ? taskParam->surfacePerBT : uSurfaceIndex; } uSurfaceNumber += kernelParam->numSurfaces; taskParam->surfacePerBT = taskParam->surfacePerBT < uSurfaceNumber ? uSurfaceNumber : taskParam->surfacePerBT; // 26Z must be media object because by default it uses thread dependency mask // if there is no thread payload and dependency is not WAVEFRONT26Z, check if walker can be used if ( kernelParam->payloadSize == 0) { //per-kernel thread space is avaiable, and check it at first if((kernelParam->kernelThreadSpaceParam.threadSpaceWidth != 0) && (kernelParam->kernelThreadSpaceParam.patternType != CM_WAVEFRONT26Z) && (kernelParam->kernelThreadSpaceParam.patternType != CM_WAVEFRONT26ZI) && (kernelParam->kernelThreadSpaceParam.threadCoordinates == nullptr)) { kernelParam->walkerParams.cmWalkerEnable = true; kernelParam->walkerParams.groupIdLoopSelect = execParam->mediaWalkerGroupSelect; } else if (kernelParam->kernelThreadSpaceParam.threadSpaceWidth == 0) { //Check per-task thread space setting if (state->taskParam->threadCoordinates) { if (state->taskParam->threadCoordinates[krn] == nullptr) { kernelParam->walkerParams.cmWalkerEnable = true; kernelParam->walkerParams.groupIdLoopSelect = execParam->mediaWalkerGroupSelect; } } else { kernelParam->walkerParams.cmWalkerEnable = true; kernelParam->walkerParams.groupIdLoopSelect = execParam->mediaWalkerGroupSelect; } } } //Media walker mode will be disabled if any kernel need use media object, we don't support mixed working modes state->walkerParams.CmWalkerEnable &= kernelParam->walkerParams.cmWalkerEnable; if (!state->walkerParams.CmWalkerEnable) { taskParam->batchBufferSize += kernelParam->numThreads * (hdrSize + MOS_MAX(kernelParam->payloadSize, 4)); } totalThreads += kernelParam->numThreads; } taskParam->batchBufferSize += CM_EXTRA_BB_SPACE; if (state->cmHalInterface->IsScoreboardParamNeeded()) { scoreboardParams = &state->scoreboardParams; scoreboardParams->ScoreboardMask = 0; scoreboardParams->ScoreboardType = nonstallingScoreboardEnable; // set VFE scoreboarding information from union of kernel dependency vectors MOS_ZeroMemory(&vfeDependencyInfo, sizeof(CM_HAL_DEPENDENCY)); for (krn = 0; krn < execParam->numKernels; krn++) { kernelParam = execParam->kernels[krn]; kernelTSParam = &kernelParam->kernelThreadSpaceParam; // calculate union dependency vector of all kernels with dependency if (kernelTSParam->dependencyInfo.count || kernelTSParam->dependencyVectorsValid) { if (vfeDependencyInfo.count == 0) { if (kernelTSParam->dependencyInfo.count) { MOS_SecureMemcpy(&vfeDependencyInfo, sizeof(CM_HAL_DEPENDENCY), &kernelTSParam->dependencyInfo, sizeof(CM_HAL_DEPENDENCY)); } else if (kernelTSParam->dependencyVectorsValid) { MOS_SecureMemcpy(&vfeDependencyInfo, sizeof(CM_HAL_DEPENDENCY), &kernelTSParam->dependencyVectors, sizeof(CM_HAL_DEPENDENCY)); } kernelTSParam->globalDependencyMask = (1 << vfeDependencyInfo.count) - 1; } else { uint32_t count = 0; CM_HAL_DEPENDENCY dependencyInfo; if (kernelTSParam->dependencyVectorsValid) { count = kernelTSParam->dependencyVectors.count; MOS_SecureMemcpy(&dependencyInfo.deltaX, sizeof(int32_t) * count, &kernelTSParam->dependencyVectors.deltaX, sizeof(int32_t) * count); MOS_SecureMemcpy(&dependencyInfo.deltaY, sizeof(int32_t) * count, &kernelTSParam->dependencyVectors.deltaY, sizeof(int32_t) * count); } else { count = kernelTSParam->dependencyInfo.count; MOS_SecureMemcpy(&dependencyInfo.deltaX, sizeof(int32_t) * count, &kernelTSParam->dependencyInfo.deltaX, sizeof(int32_t) * count); MOS_SecureMemcpy(&dependencyInfo.deltaY, sizeof(int32_t) * count, &kernelTSParam->dependencyInfo.deltaY, sizeof(int32_t) * count); } for (j = 0; j < count; ++j) { for (k = 0; k < vfeDependencyInfo.count; ++k) { if ((dependencyInfo.deltaX[j] == vfeDependencyInfo.deltaX[k]) && (dependencyInfo.deltaY[j] == vfeDependencyInfo.deltaY[k])) { CM_HAL_SETBIT(kernelTSParam->globalDependencyMask, k); break; } } if (k == vfeDependencyInfo.count) { vfeDependencyInfo.deltaX[vfeDependencyInfo.count] = dependencyInfo.deltaX[j]; vfeDependencyInfo.deltaY[vfeDependencyInfo.count] = dependencyInfo.deltaY[j]; CM_HAL_SETBIT(kernelTSParam->globalDependencyMask, vfeDependencyInfo.count); vfeDependencyInfo.count++; } } } } reuseBBUpdateMask |= kernelTSParam->reuseBBUpdateMask; } if (vfeDependencyInfo.count > CM_HAL_MAX_DEPENDENCY_COUNT) { eStatus = MOS_STATUS_INVALID_PARAMETER; CM_ASSERTMESSAGE("Union of kernel dependencies exceeds max dependency count (8)"); goto finish; } scoreboardParams->ScoreboardMask = (uint8_t)vfeDependencyInfo.count; for (i = 0; i < scoreboardParams->ScoreboardMask; ++i) { scoreboardParams->ScoreboardDelta[i].x = vfeDependencyInfo.deltaX[i]; scoreboardParams->ScoreboardDelta[i].y = vfeDependencyInfo.deltaY[i]; } //If no dependency defined in kernel data, then check per-task thread space setting if (scoreboardParams->ScoreboardMask == 0) { if (taskParam->dependencyVectorsValid) { scoreboardParams->ScoreboardMask = (uint8_t)taskParam->dependencyVectors.count; for (uint32_t i = 0; i < scoreboardParams->ScoreboardMask; ++i) { scoreboardParams->ScoreboardDelta[i].x = taskParam->dependencyVectors.deltaX[i]; scoreboardParams->ScoreboardDelta[i].y = taskParam->dependencyVectors.deltaY[i]; } } else { switch (taskParam->dependencyPattern) { case CM_NONE_DEPENDENCY: break; case CM_VERTICAL_WAVE: scoreboardParams->ScoreboardMask = 1; scoreboardParams->ScoreboardDelta[0].x = 0xF; // -1 in uint8_t:4 scoreboardParams->ScoreboardDelta[0].y = 0; break; case CM_HORIZONTAL_WAVE: scoreboardParams->ScoreboardMask = 1; scoreboardParams->ScoreboardDelta[0].x = 0; scoreboardParams->ScoreboardDelta[0].y = 0xF; // -1 in uint8_t:4 break; case CM_WAVEFRONT: scoreboardParams->ScoreboardMask = 3; scoreboardParams->ScoreboardDelta[0].x = 0xF; // -1 in uint8_t:4 scoreboardParams->ScoreboardDelta[0].y = 0; scoreboardParams->ScoreboardDelta[1].x = 0xF; // -1 in uint8_t:4 scoreboardParams->ScoreboardDelta[1].y = 0xF; // -1 in uint8_t:4 scoreboardParams->ScoreboardDelta[2].x = 0; scoreboardParams->ScoreboardDelta[2].y = 0xF; // -1 in uint8_t:4 break; case CM_WAVEFRONT26: scoreboardParams->ScoreboardMask = 4; scoreboardParams->ScoreboardDelta[0].x = 0xF; // -1 in uint8_t:4 scoreboardParams->ScoreboardDelta[0].y = 0; scoreboardParams->ScoreboardDelta[1].x = 0xF; // -1 in uint8_t:4 scoreboardParams->ScoreboardDelta[1].y = 0xF; // -1 in uint8_t:4 scoreboardParams->ScoreboardDelta[2].x = 0; scoreboardParams->ScoreboardDelta[2].y = 0xF; // -1 in uint8_t:4 scoreboardParams->ScoreboardDelta[3].x = 1; scoreboardParams->ScoreboardDelta[3].y = 0xF; // -1 in uint8_t:4 break; case CM_WAVEFRONT26Z: case CM_WAVEFRONT26ZIG: scoreboardParams->ScoreboardMask = 5; scoreboardParams->ScoreboardDelta[0].x = 0xF; // -1 in uint8_t:4 scoreboardParams->ScoreboardDelta[0].y = 1; scoreboardParams->ScoreboardDelta[1].x = 0xF; // -1 in uint8_t:4 scoreboardParams->ScoreboardDelta[1].y = 0; scoreboardParams->ScoreboardDelta[2].x = 0xF; // -1 in uint8_t:4 scoreboardParams->ScoreboardDelta[2].y = 0xF; // -1 in uint8_t:4 scoreboardParams->ScoreboardDelta[3].x = 0; scoreboardParams->ScoreboardDelta[3].y = 0xF; // -1 in uint8_t:4 scoreboardParams->ScoreboardDelta[4].x = 1; scoreboardParams->ScoreboardDelta[4].y = 0xF; // -1 in uint8_t:4 break; case CM_WAVEFRONT26ZI: scoreboardParams->ScoreboardMask = 7; scoreboardParams->ScoreboardDelta[0].x = 0xF; // -1 in uint8_t:4 scoreboardParams->ScoreboardDelta[0].y = 1; scoreboardParams->ScoreboardDelta[1].x = 0xE; // -2 scoreboardParams->ScoreboardDelta[1].y = 0; scoreboardParams->ScoreboardDelta[2].x = 0xF; // -1 in uint8_t:4 scoreboardParams->ScoreboardDelta[2].y = 0; scoreboardParams->ScoreboardDelta[3].x = 0xF; // -1 in uint8_t:4 scoreboardParams->ScoreboardDelta[3].y = 0xF; // -1 in uint8_t:4 scoreboardParams->ScoreboardDelta[4].x = 0; scoreboardParams->ScoreboardDelta[4].y = 0xF; // -1 in uint8_t:4 scoreboardParams->ScoreboardDelta[5].x = 1; scoreboardParams->ScoreboardDelta[5].y = 0xF; // -1 in uint8_t:4 scoreboardParams->ScoreboardDelta[6].x = 1; scoreboardParams->ScoreboardDelta[6].y = 0; break; case CM_WAVEFRONT26X: scoreboardParams->ScoreboardMask = 7; scoreboardParams->ScoreboardDelta[0].x = 0xF; scoreboardParams->ScoreboardDelta[0].y = 3; scoreboardParams->ScoreboardDelta[1].x = 0xF; scoreboardParams->ScoreboardDelta[1].y = 1; scoreboardParams->ScoreboardDelta[2].x = 0xF; scoreboardParams->ScoreboardDelta[2].y = 0xF; scoreboardParams->ScoreboardDelta[3].x = 0; scoreboardParams->ScoreboardDelta[3].y = 0xF; scoreboardParams->ScoreboardDelta[4].x = 0; scoreboardParams->ScoreboardDelta[4].y = 0xE; scoreboardParams->ScoreboardDelta[5].x = 0; scoreboardParams->ScoreboardDelta[5].y = 0xD; scoreboardParams->ScoreboardDelta[6].x = 1; scoreboardParams->ScoreboardDelta[6].y = 0xD; break; default: taskParam->dependencyPattern = CM_NONE_DEPENDENCY; break; } } } } //Set size of surface binding table size CM_SURFACE_BTI_INFO surfBTIInfo; state->cmHalInterface->GetHwSurfaceBTIInfo(&surfBTIInfo); taskParam->surfacePerBT += surfBTIInfo.normalSurfaceStart ; // add one if kernel debugger is enabled if (execParam->kernelDebugEnabled) { taskParam->surfacePerBT += CM_RESERVED_SURFACE_NUMBER_FOR_KERNEL_DEBUG; } //If global surface is used and current surface bt size less than the max index of reserved surfaces //use set it as max bti size if ((execParam->globalSurfaceUsed) && (taskParam->surfacePerBT < surfBTIInfo.reservedSurfaceEnd)) { taskParam->surfacePerBT = CM_MAX_STATIC_SURFACE_STATES_PER_BT; } //Make sure surfacePerBT do not exceed CM_MAX_STATIC_SURFACE_STATES_PER_BT taskParam->surfacePerBT = MOS_MIN(CM_MAX_STATIC_SURFACE_STATES_PER_BT, taskParam->surfacePerBT); if( taskParam->dependencyMasks ) { for (krn = 0; krn < execParam->numKernels; krn++) { kernelParam = execParam->kernels[krn]; dependencyMask = taskParam->dependencyMasks[krn]; if( dependencyMask ) { for( i = 0; i < kernelParam->numThreads; ++i ) { reuseBBUpdateMask |= dependencyMask[i].resetMask; } } } } CM_HAL_CHECKBIT_IS_SET(bitIsSet, reuseBBUpdateMask, CM_NO_BATCH_BUFFER_REUSE_BIT_POS); if( bitIsSet || reuseBBUpdateMask == 0 ) { taskParam->reuseBBUpdateMask = 0; } else { taskParam->reuseBBUpdateMask = 1; } threadArgExists = HalCm_GetTaskHasThreadArg(execParam->kernels, execParam->numKernels); // For media object with thread arg, only support up to CM_MAX_USER_THREADS (512*512) threads // otherwise can support up to 262144 media object commands in batch buffer if (!state->walkerParams.CmWalkerEnable) { if (!threadArgExists) { if(totalThreads > CM_MAX_USER_THREADS_NO_THREADARG) { eStatus = MOS_STATUS_INVALID_PARAMETER; CM_ASSERTMESSAGE( "Total task threads '%d' exceeds max allowed threads '%d'", totalThreads, CM_MAX_USER_THREADS_NO_THREADARG); goto finish; } } else { if (totalThreads > CM_MAX_USER_THREADS) { eStatus = MOS_STATUS_INVALID_PARAMETER; CM_ASSERTMESSAGE( "Total task threads '%d' exceeds max allowed threads '%d'", totalThreads, CM_MAX_USER_THREADS); goto finish; } } } taskParam->queueOption = execParam->queueOption; finish: return eStatus; } //*----------------------------------------------------------------------------- //| Purpose: Parse the Kernel and populate the Task Param structure //| Return: Result of the operation //*----------------------------------------------------------------------------- MOS_STATUS HalCm_ParseGroupTask( PCM_HAL_STATE state, // [in] Pointer to HAL CM state PCM_HAL_EXEC_GROUP_TASK_PARAM execGroupParam) // [in] Pointer to Exec Task Param { PCM_HAL_TASK_PARAM taskParam = state->taskParam; MOS_STATUS eStatus = MOS_STATUS_SUCCESS; PCM_HAL_KERNEL_PARAM kernelParam = nullptr; uint32_t uSurfaceIndex; taskParam->surfEntryInfoArrays = execGroupParam->surEntryInfoArrays; //GT-PIN taskParam->batchBufferSize = 0; taskParam->kernelDebugEnabled = (uint32_t)execGroupParam->kernelDebugEnabled; taskParam->numKernels = execGroupParam->numKernels; taskParam->syncBitmap = execGroupParam->syncBitmap; taskParam->conditionalEndBitmap = execGroupParam->conditionalEndBitmap; MOS_SecureMemcpy(taskParam->conditionalEndInfo, sizeof(taskParam->conditionalEndInfo), execGroupParam->conditionalEndInfo, sizeof(execGroupParam->conditionalEndInfo)); taskParam->taskConfig = execGroupParam->taskConfig; MOS_SecureMemcpy(taskParam->krnExecCfg, sizeof(taskParam->krnExecCfg), execGroupParam->krnExecCfg, sizeof(execGroupParam->krnExecCfg)); for (uint32_t krn = 0; krn < execGroupParam->numKernels; krn ++) { kernelParam = execGroupParam->kernels[krn]; PCM_INDIRECT_SURFACE_INFO indirectSurfaceInfo = kernelParam->indirectDataParam.surfaceInfo; uint32_t uSurfaceNumber = 0; if (kernelParam->indirectDataParam.surfaceCount) { uSurfaceIndex = 0; for (uint32_t i = 0; i < kernelParam->indirectDataParam.surfaceCount; i++) { uSurfaceIndex = (indirectSurfaceInfo + i)->bindingTableIndex > uSurfaceIndex ? (indirectSurfaceInfo + i)->bindingTableIndex : uSurfaceIndex; uSurfaceNumber++; } taskParam->surfacePerBT = taskParam->surfacePerBT > uSurfaceIndex ? taskParam->surfacePerBT : uSurfaceIndex; } uSurfaceNumber += kernelParam->numSurfaces; taskParam->surfacePerBT = taskParam->surfacePerBT < uSurfaceNumber ? uSurfaceNumber : taskParam->surfacePerBT; } CM_SURFACE_BTI_INFO surfBTIInfo; state->cmHalInterface->GetHwSurfaceBTIInfo(&surfBTIInfo); taskParam->surfacePerBT += surfBTIInfo.normalSurfaceStart ; // add one if kernel debugger is enabled if (execGroupParam->kernelDebugEnabled) { taskParam->surfacePerBT += CM_RESERVED_SURFACE_NUMBER_FOR_KERNEL_DEBUG; } //If global surface is used and current surface bt size less than the max index of reserved surfaces //use set it as max bti size if ((execGroupParam->globalSurfaceUsed) && (taskParam->surfacePerBT < surfBTIInfo.reservedSurfaceEnd)) { taskParam->surfacePerBT = CM_MAX_STATIC_SURFACE_STATES_PER_BT; } //Make sure surfacePerBT do not exceed CM_MAX_STATIC_SURFACE_STATES_PER_BT taskParam->surfacePerBT = MOS_MIN(CM_MAX_STATIC_SURFACE_STATES_PER_BT, taskParam->surfacePerBT); taskParam->queueOption = execGroupParam->queueOption; taskParam->mosVeHintParams = execGroupParam->mosVeHintParams; return eStatus; } //*----------------------------------------------------------------------------- //| Purpose: Parse the Kernel and populate the Hints Task Param structure //| Return: Result of the operation //*----------------------------------------------------------------------------- MOS_STATUS HalCm_ParseHintsTask( PCM_HAL_STATE state, // [in] Pointer to HAL CM state PCM_HAL_EXEC_HINTS_TASK_PARAM execHintsParam) { MOS_STATUS eStatus; PCM_HAL_TASK_PARAM taskParam; PCM_HAL_KERNEL_PARAM kernelParam; uint32_t hdrSize; uint32_t totalThreads; uint32_t krn; uint32_t curbeOffset; PMHW_VFE_SCOREBOARD scoreboardParams; uint32_t hasThreadArg; bool nonstallingScoreboardEnable; bool bitIsSet; uint8_t reuseBBUpdateMask; bool threadArgExists; eStatus = MOS_STATUS_SUCCESS; krn = 0; taskParam = state->taskParam; nonstallingScoreboardEnable = true; bitIsSet = false; curbeOffset = 0; hasThreadArg = 0; totalThreads = 0; reuseBBUpdateMask = 0; threadArgExists = false; hdrSize = state->renderHal->pHwSizes->dwSizeMediaObjectHeaderCmd; scoreboardParams = &state->scoreboardParams; for( krn = 0; krn < execHintsParam->numKernels; ++krn ) { if ((execHintsParam->kernels[krn] == nullptr) || (execHintsParam->kernelSizes[krn] == 0)) { eStatus = MOS_STATUS_INVALID_PARAMETER; CM_ASSERTMESSAGE("Invalid Kernel data"); goto finish; } // Parse the kernel Param kernelParam = execHintsParam->kernels[krn]; // if any kernel disables non-stalling, the non-stalling will be disabled nonstallingScoreboardEnable &= (kernelParam->cmFlags & CM_KERNEL_FLAGS_NONSTALLING_SCOREBOARD) ? true : false; if (!state->walkerParams.CmWalkerEnable) { taskParam->batchBufferSize += kernelParam->numThreads * (hdrSize + MOS_MAX(kernelParam->payloadSize, 4)); } totalThreads += kernelParam->numThreads; reuseBBUpdateMask |= kernelParam->kernelThreadSpaceParam.reuseBBUpdateMask; } CM_HAL_CHECKBIT_IS_SET(bitIsSet, reuseBBUpdateMask, CM_NO_BATCH_BUFFER_REUSE_BIT_POS); if( bitIsSet || reuseBBUpdateMask == 0 ) { taskParam->reuseBBUpdateMask = 0; } else { taskParam->reuseBBUpdateMask = 1; } taskParam->batchBufferSize += CM_EXTRA_BB_SPACE; scoreboardParams->ScoreboardType = nonstallingScoreboardEnable; threadArgExists = HalCm_GetTaskHasThreadArg(execHintsParam->kernels, execHintsParam->numKernels); if (!state->walkerParams.CmWalkerEnable) { if (!threadArgExists) { if(totalThreads > CM_MAX_USER_THREADS_NO_THREADARG) { eStatus = MOS_STATUS_INVALID_PARAMETER; CM_ASSERTMESSAGE( "Total task threads '%d' exceeds max allowed threads '%d'", totalThreads, CM_MAX_USER_THREADS_NO_THREADARG); goto finish; } } else { if (totalThreads > CM_MAX_USER_THREADS) { eStatus = MOS_STATUS_INVALID_PARAMETER; CM_ASSERTMESSAGE( "Total task threads '%d' exceeds max allowed threads '%d'", totalThreads, CM_MAX_USER_THREADS); goto finish; } } } taskParam->queueOption = execHintsParam->queueOption; finish: return eStatus; } /* ** check to see if kernel entry is flaged as free or it is null ** used for combining */ bool bIsFree( PRENDERHAL_KRN_ALLOCATION kAlloc ) { if (kAlloc== nullptr) { return false; } else { if (kAlloc->dwFlags != RENDERHAL_KERNEL_ALLOCATION_FREE) { return false; } } return true; } /* ** local used supporting function ** setup correct values according to input and copy kernelBinary as needed */ void CmLoadKernel(PCM_HAL_STATE state, PRENDERHAL_STATE_HEAP stateHeap, PRENDERHAL_KRN_ALLOCATION kernelAllocation, uint32_t sync, uint32_t count, PRENDERHAL_KERNEL_PARAM parameters, PCM_HAL_KERNEL_PARAM kernelParam, MHW_KERNEL_PARAM *mhwKernelParam, bool isCloneEntry) { UNUSED(state); if (mhwKernelParam) { kernelAllocation->iKID = -1; kernelAllocation->iKUID = mhwKernelParam->iKUID; kernelAllocation->iKCID = mhwKernelParam->iKCID; kernelAllocation->dwSync = sync; kernelAllocation->dwCount = count & 0xFFFFFFFF; // 28 bits kernelAllocation->dwFlags = RENDERHAL_KERNEL_ALLOCATION_USED; kernelAllocation->Params = *parameters; kernelAllocation->pMhwKernelParam = mhwKernelParam; if (!isCloneEntry) { // Copy kernel data // Copy MovInstruction First MOS_SecureMemcpy(stateHeap->pIshBuffer + kernelAllocation->dwOffset, kernelParam->movInsDataSize, kernelParam->movInsData, kernelParam->movInsDataSize); // Copy Cm Kernel Binary MOS_SecureMemcpy(stateHeap->pIshBuffer + kernelAllocation->dwOffset + kernelParam->movInsDataSize, kernelParam->kernelBinarySize - kernelParam->movInsDataSize, kernelParam->kernelBinary, kernelParam->kernelBinarySize - kernelParam->movInsDataSize); // Padding bytes dummy instructions after kernel binary to resolve page fault issue MOS_ZeroMemory(stateHeap->pIshBuffer + kernelAllocation->dwOffset + kernelParam->kernelBinarySize, CM_KERNEL_BINARY_PADDING_SIZE); } } else { kernelAllocation->iKID = -1; kernelAllocation->iKUID = -1; kernelAllocation->iKCID = -1; kernelAllocation->dwSync = 0; FrameTrackerTokenFlat_Clear(&kernelAllocation->trackerToken); kernelAllocation->dwCount = 0; kernelAllocation->dwFlags = RENDERHAL_KERNEL_ALLOCATION_FREE; kernelAllocation->pMhwKernelParam = nullptr; kernelAllocation->cloneKernelParams.cloneKernelID = -1; kernelAllocation->cloneKernelParams.isClone = false; kernelAllocation->cloneKernelParams.isHeadKernel = false; kernelAllocation->cloneKernelParams.kernelBinaryAllocID = -1; kernelAllocation->cloneKernelParams.referenceCount = 0; } } /* ** local used supporting function ** Try to find free entry which is big enough to load kernel binary ** If we cannot find one, then return fail, so we will delete more entries */ int32_t CmSearchFreeSlotSize(PCM_HAL_STATE state, MHW_KERNEL_PARAM *mhwKernelParam, bool isCloneEntry) { PRENDERHAL_STATE_HEAP stateHeap; PRENDERHAL_KRN_ALLOCATION kernelAllocation; int32_t kernelAllocationID; int32_t returnVal = -1; int32_t neededSize; stateHeap = state->renderHal->pStateHeap; kernelAllocation = stateHeap->pKernelAllocation; if (isCloneEntry) { neededSize = CM_64BYTE; } else { neededSize = mhwKernelParam->iSize; } for (kernelAllocationID = 0; kernelAllocationID < state->kernelNumInGsh; kernelAllocationID++, kernelAllocation++) { if(kernelAllocation->dwFlags == RENDERHAL_KERNEL_ALLOCATION_FREE) { if(state->totalKernelSize[kernelAllocationID] >= neededSize) { // found free slot which is big enough return kernelAllocationID; } } } // not found return returnVal; } //*----------------------------------------------------------------------------- //| Purpose: Updates the clone entries' head kernel binary allocation IDs //| Function is called after kernel allocations are shifted due to combining neighboring free entries //| Return: Result of the operation //*----------------------------------------------------------------------------- void HalCm_UpdateCloneKernel(PCM_HAL_STATE state, uint32_t shiftPoint, CM_SHIFT_DIRECTION shiftDirection, uint32_t shiftFactor) { PRENDERHAL_STATE_HEAP stateHeap; PRENDERHAL_KRN_ALLOCATION kernelAllocation; int32_t allocationID; stateHeap = state->renderHal->pStateHeap; kernelAllocation = stateHeap->pKernelAllocation; for (allocationID = 0; allocationID < state->kernelNumInGsh; allocationID++, kernelAllocation++) { kernelAllocation = &(stateHeap->pKernelAllocation[allocationID]); if (kernelAllocation->cloneKernelParams.isClone && ((kernelAllocation->cloneKernelParams.kernelBinaryAllocID) > (int32_t)shiftPoint)) { if (shiftDirection == CM_SHIFT_LEFT) { kernelAllocation->cloneKernelParams.kernelBinaryAllocID = kernelAllocation->cloneKernelParams.kernelBinaryAllocID + shiftFactor; } else { kernelAllocation->cloneKernelParams.kernelBinaryAllocID = kernelAllocation->cloneKernelParams.kernelBinaryAllocID - shiftFactor; } } } } /* ** local used supporting function ** We found free slot and load kernel to this slot. There are 3 cases (see code) */ int32_t CmAddCurrentKernelToFreeSlot(PCM_HAL_STATE state, int32_t slot, PRENDERHAL_KERNEL_PARAM parameters, PCM_HAL_KERNEL_PARAM kernelParam, MHW_KERNEL_PARAM *mhwKernelParam, CM_CLONE_TYPE cloneType, int32_t headKernelAllocationID) { PRENDERHAL_STATE_HEAP stateHeap; PRENDERHAL_KRN_ALLOCATION kernelAllocation, pKernelAllocationN; int32_t hr = CM_SUCCESS; int32_t i; int32_t totalSize, tmpSize, dwOffset, neededSize; bool adjust, isCloneEntry, isHeadKernel, isCloneAsHead, adjustHeadKernelID; uint32_t tag; stateHeap = state->renderHal->pStateHeap; kernelAllocation = stateHeap->pKernelAllocation; adjustHeadKernelID = false; switch (cloneType) { case CM_CLONE_ENTRY: { neededSize = CM_64BYTE; isCloneEntry = true; isHeadKernel = false; isCloneAsHead = false; } break; case CM_HEAD_KERNEL: { neededSize = mhwKernelParam->iSize; isHeadKernel = true; isCloneEntry = false; isCloneAsHead = false; } break; case CM_CLONE_AS_HEAD_KERNEL: { neededSize = mhwKernelParam->iSize; isHeadKernel = true; isCloneEntry = false; isCloneAsHead = true; } break; case CM_NO_CLONE: { neededSize = mhwKernelParam->iSize; isCloneEntry = false; isHeadKernel = false; isCloneAsHead = false; } break; default: { hr = CM_FAILURE; goto finish; } } // to check if we have perfect size match if(stateHeap->pKernelAllocation[slot].iSize == neededSize) { adjust = false; } else { adjust = true; } if ((state->kernelNumInGsh < state->cmDeviceParam.maxGshKernelEntries) && adjust) { // we have extra entry to add // add new entry and pump index down below int32_t lastKernel = state->kernelNumInGsh - 1; for(i = lastKernel; i>slot; i--) { kernelAllocation = &stateHeap->pKernelAllocation[i]; pKernelAllocationN = &stateHeap->pKernelAllocation[i+1]; *pKernelAllocationN = *kernelAllocation; state->totalKernelSize[i+1] = state->totalKernelSize[i]; } if (lastKernel > slot) { // update the headKernelAllocationID if it was shifted if (headKernelAllocationID > slot) { headKernelAllocationID++; adjustHeadKernelID = true; } } totalSize = state->totalKernelSize[slot]; tmpSize = neededSize; dwOffset = stateHeap->pKernelAllocation[slot].dwOffset; // now add new one kernelAllocation = &stateHeap->pKernelAllocation[slot]; if(state->cbbEnabled) { tag = state->osInterface->pfnGetGpuStatusTag(state->osInterface, state->osInterface->CurrentGpuContextOrdinal); } else { tag = stateHeap->dwNextTag; } CmLoadKernel(state, stateHeap, kernelAllocation, tag, stateHeap->dwAccessCounter, parameters, kernelParam, mhwKernelParam, isCloneEntry); stateHeap->dwAccessCounter++; kernelAllocation->iSize = tmpSize; state->totalKernelSize[slot] = MOS_ALIGN_CEIL(tmpSize, 64); // insert a new slot which is free with rest tmpSize = MOS_ALIGN_CEIL(tmpSize, 64); // HW required 64 byte align kernelAllocation = &stateHeap->pKernelAllocation[slot+1]; CmLoadKernel(state, stateHeap, kernelAllocation, 0, 0, parameters, kernelParam, nullptr, isCloneEntry); kernelAllocation->dwOffset = dwOffset+tmpSize; kernelAllocation->iSize = 0; state->totalKernelSize[slot+1] = totalSize - tmpSize; // added one more entry state->kernelNumInGsh++; kernelAllocation = &stateHeap->pKernelAllocation[slot]; if (isCloneEntry) { if (!stateHeap->pKernelAllocation[headKernelAllocationID].cloneKernelParams.isHeadKernel) { // ERROR thought kernel with allocation ID, headKernelAllocationID, was a head kernel, but it's not hr = CM_FAILURE; goto finish; } kernelAllocation->cloneKernelParams.dwOffsetForAllocID = dwOffset; kernelAllocation->dwOffset = stateHeap->pKernelAllocation[headKernelAllocationID].dwOffset; kernelAllocation->cloneKernelParams.isClone = true; kernelAllocation->cloneKernelParams.kernelBinaryAllocID = headKernelAllocationID; kernelAllocation->cloneKernelParams.cloneKernelID = stateHeap->pKernelAllocation[headKernelAllocationID].iKUID; stateHeap->pKernelAllocation[headKernelAllocationID].cloneKernelParams.referenceCount = stateHeap->pKernelAllocation[headKernelAllocationID].cloneKernelParams.referenceCount + 1; // update head kernel's count after updating the clone entry's count so that clone will be selected for deletion first stateHeap->pKernelAllocation[headKernelAllocationID].dwCount = stateHeap->dwAccessCounter++; } else { kernelAllocation->dwOffset = dwOffset; if (isHeadKernel) { kernelAllocation->cloneKernelParams.isHeadKernel = true; if (isCloneAsHead) { kernelAllocation->cloneKernelParams.cloneKernelID = kernelParam->clonedKernelParam.kernelID; } } } if (lastKernel > slot) { HalCm_UpdateCloneKernel(state, slot, CM_SHIFT_LEFT, 1); if (isCloneEntry && adjustHeadKernelID) { // if clone entry and already adjusted head kernel ID, then adjusted again in HalCm_UpdateCloneKernel, need to do only once kernelAllocation->cloneKernelParams.kernelBinaryAllocID = kernelAllocation->cloneKernelParams.kernelBinaryAllocID - 1; } } } else if (state->kernelNumInGsh < state->cmDeviceParam.maxGshKernelEntries) { // no need to create a new entry since we have the same size kernelAllocation = &stateHeap->pKernelAllocation[slot]; if(state->cbbEnabled) { tag = state->osInterface->pfnGetGpuStatusTag(state->osInterface, state->osInterface->CurrentGpuContextOrdinal); } else { tag = stateHeap->dwNextTag; } CmLoadKernel(state, stateHeap, kernelAllocation, tag, stateHeap->dwAccessCounter, parameters, kernelParam, mhwKernelParam, isCloneEntry); stateHeap->dwAccessCounter++; // no change for kernelAllocation->dwOffset kernelAllocation->iSize = neededSize; state->totalKernelSize[slot] = MOS_ALIGN_CEIL(mhwKernelParam->iSize, 64); if (isCloneEntry) { if (!stateHeap->pKernelAllocation[headKernelAllocationID].cloneKernelParams.isHeadKernel) { // ERROR thought kernel with allocation ID, headKernelAllocationID, was a head kernel, but it's not hr = CM_FAILURE; goto finish; } kernelAllocation->cloneKernelParams.dwOffsetForAllocID = kernelAllocation->dwOffset; kernelAllocation->dwOffset = stateHeap->pKernelAllocation[headKernelAllocationID].dwOffset; kernelAllocation->cloneKernelParams.isClone = true; kernelAllocation->cloneKernelParams.kernelBinaryAllocID = headKernelAllocationID; kernelAllocation->cloneKernelParams.cloneKernelID = stateHeap->pKernelAllocation[headKernelAllocationID].iKUID; stateHeap->pKernelAllocation[headKernelAllocationID].cloneKernelParams.referenceCount = stateHeap->pKernelAllocation[headKernelAllocationID].cloneKernelParams.referenceCount + 1; // update head kernel's count after updating the clone entry's count so that clone will be selected for deletion first stateHeap->pKernelAllocation[headKernelAllocationID].dwCount = stateHeap->dwAccessCounter++; } else if (isHeadKernel) { kernelAllocation->cloneKernelParams.isHeadKernel = true; if (isCloneAsHead) { kernelAllocation->cloneKernelParams.cloneKernelID = kernelParam->clonedKernelParam.kernelID; } } } else { // all slots are used, but we have one free which is big enough // we may have fragmentation, but code is the same as above case kernelAllocation = &stateHeap->pKernelAllocation[slot]; if(state->cbbEnabled) { tag = state->osInterface->pfnGetGpuStatusTag(state->osInterface, state->osInterface->CurrentGpuContextOrdinal); } else { tag = stateHeap->dwNextTag; } CmLoadKernel(state, stateHeap, kernelAllocation, tag, stateHeap->dwAccessCounter, parameters, kernelParam, mhwKernelParam, isCloneEntry); stateHeap->dwAccessCounter++; // kernelAllocation->iTotalSize is not changed, but we have smaller actual size // no change for kernelAllocation->dwOffset kernelAllocation->iSize = neededSize; if (isCloneEntry) { if (!stateHeap->pKernelAllocation[headKernelAllocationID].cloneKernelParams.isHeadKernel) { // ERROR thought kernel with allocation ID, headKernelAllocationID, was a head kernel, but it's not hr = CM_FAILURE; goto finish; } kernelAllocation->cloneKernelParams.dwOffsetForAllocID = kernelAllocation->dwOffset; kernelAllocation->dwOffset = stateHeap->pKernelAllocation[headKernelAllocationID].dwOffset; kernelAllocation->cloneKernelParams.isClone = true; kernelAllocation->cloneKernelParams.kernelBinaryAllocID = headKernelAllocationID; kernelAllocation->cloneKernelParams.cloneKernelID = stateHeap->pKernelAllocation[headKernelAllocationID].iKUID; stateHeap->pKernelAllocation[headKernelAllocationID].cloneKernelParams.referenceCount = stateHeap->pKernelAllocation[headKernelAllocationID].cloneKernelParams.referenceCount + 1; // update head kernel's count after updating the clone entry's count so that clone will be selected for deletion first stateHeap->pKernelAllocation[headKernelAllocationID].dwCount = stateHeap->dwAccessCounter++; } else if (isHeadKernel) { kernelAllocation->cloneKernelParams.isHeadKernel = true; if (isCloneAsHead) { kernelAllocation->cloneKernelParams.cloneKernelID = kernelParam->clonedKernelParam.kernelID; } } } finish: return hr; } /*---------------------------------------------------------------------------- | Name : HalCm_UnLoadKernel ( Replace RenderHal_UnloadKernel) \---------------------------------------------------------------------------*/ int32_t HalCm_UnloadKernel( PCM_HAL_STATE state, PRENDERHAL_KRN_ALLOCATION kernelAllocation) { PRENDERHAL_INTERFACE renderHal = state->renderHal; PRENDERHAL_STATE_HEAP stateHeap; int32_t hr; //--------------------------------------- CM_CHK_NULL_GOTOFINISH_CMERROR(renderHal); CM_CHK_NULL_GOTOFINISH_CMERROR(renderHal->pStateHeap); CM_CHK_NULL_GOTOFINISH_CMERROR(kernelAllocation); //--------------------------------------- hr = CM_FAILURE; stateHeap = renderHal->pStateHeap; if (kernelAllocation->dwFlags == RENDERHAL_KERNEL_ALLOCATION_FREE) { goto finish; } CM_CHK_CMSTATUS_GOTOFINISH(state->pfnSyncKernel(state, kernelAllocation->dwSync)); // Unload kernel if (kernelAllocation->pMhwKernelParam) { kernelAllocation->pMhwKernelParam->bLoaded = 0; } if (kernelAllocation->cloneKernelParams.isClone) { if (stateHeap->pKernelAllocation[kernelAllocation->cloneKernelParams.kernelBinaryAllocID].cloneKernelParams.isHeadKernel) { if ((stateHeap->pKernelAllocation[kernelAllocation->cloneKernelParams.kernelBinaryAllocID].cloneKernelParams.referenceCount) <= 0) { // ERROR hr = CM_FAILURE; goto finish; } } else { // ERROR hr = CM_FAILURE; goto finish; } stateHeap->pKernelAllocation[kernelAllocation->cloneKernelParams.kernelBinaryAllocID].cloneKernelParams.referenceCount = stateHeap->pKernelAllocation[kernelAllocation->cloneKernelParams.kernelBinaryAllocID].cloneKernelParams.referenceCount - 1; // restore the dwOffset for this allocationID kernelAllocation->dwOffset = kernelAllocation->cloneKernelParams.dwOffsetForAllocID; } else if (kernelAllocation->cloneKernelParams.isHeadKernel && kernelAllocation->cloneKernelParams.referenceCount != 0) { // ERROR, cloned kernel entries should have been selected for deletion before head kernel entry hr = CM_FAILURE; goto finish; } // Release kernel entry (Offset/size may be used for reallocation) kernelAllocation->iKID = -1; kernelAllocation->iKUID = -1; kernelAllocation->iKCID = -1; kernelAllocation->dwSync = 0; FrameTrackerTokenFlat_Clear(&kernelAllocation->trackerToken); kernelAllocation->dwFlags = RENDERHAL_KERNEL_ALLOCATION_FREE; kernelAllocation->dwCount = 0; kernelAllocation->pMhwKernelParam = nullptr; kernelAllocation->cloneKernelParams.cloneKernelID = -1; kernelAllocation->cloneKernelParams.isClone = false; kernelAllocation->cloneKernelParams.isHeadKernel = false; kernelAllocation->cloneKernelParams.kernelBinaryAllocID = -1; kernelAllocation->cloneKernelParams.referenceCount = 0; hr = CM_SUCCESS; finish: return hr; } /*---------------------------------------------------------------------------- | Name : HalCmw_TouchKernel ( Replace RenderHal_TouchKernel) \---------------------------------------------------------------------------*/ int32_t HalCm_TouchKernel( PCM_HAL_STATE state, int32_t kernelAllocationID) { int32_t hr = CM_SUCCESS; PRENDERHAL_STATE_HEAP stateHeap; PRENDERHAL_KRN_ALLOCATION kernelAllocation; PRENDERHAL_KRN_ALLOCATION headKernelAllocation; uint32_t tag; PRENDERHAL_INTERFACE renderHal = state->renderHal; PMOS_INTERFACE osInterface = state->osInterface; stateHeap = (renderHal) ? renderHal->pStateHeap : nullptr; if (stateHeap == nullptr || stateHeap->pKernelAllocation == nullptr || kernelAllocationID < 0 || kernelAllocationID >= renderHal->StateHeapSettings.iKernelCount) { hr = CM_FAILURE; goto finish; } // Update usage kernelAllocation = &(stateHeap->pKernelAllocation[kernelAllocationID]); if (kernelAllocation->dwFlags != RENDERHAL_KERNEL_ALLOCATION_FREE && kernelAllocation->dwFlags != RENDERHAL_KERNEL_ALLOCATION_LOCKED) { kernelAllocation->dwCount = stateHeap->dwAccessCounter++; } // Set sync tag, for deallocation control if(state->cbbEnabled) { tag = osInterface->pfnGetGpuStatusTag(osInterface, osInterface->CurrentGpuContextOrdinal); } else { tag = stateHeap->dwNextTag; } kernelAllocation->dwSync = tag; // if this kernel allocation is a cloned kernel, update the orig kernel sync tag and access counter if (kernelAllocation->cloneKernelParams.isClone) { headKernelAllocation = &(stateHeap->pKernelAllocation[kernelAllocation->cloneKernelParams.kernelBinaryAllocID]); if (headKernelAllocation->cloneKernelParams.referenceCount <= 0) { // ERROR hr = CM_FAILURE; goto finish; } headKernelAllocation->dwSync = tag; headKernelAllocation->dwCount = stateHeap->dwAccessCounter++; } finish: return hr; } /* ** Supporting function ** Delete oldest entry from table to free more space ** According to different cases, we will combine space with previous or next slot to get max space */ int32_t CmDeleteOldestKernel(PCM_HAL_STATE state, MHW_KERNEL_PARAM *mhwKernelParam) { PRENDERHAL_KRN_ALLOCATION kernelAllocation; PRENDERHAL_INTERFACE renderHal = state->renderHal;; PRENDERHAL_STATE_HEAP stateHeap = renderHal->pStateHeap; UNUSED(state); UNUSED(mhwKernelParam); uint32_t oldest = 0; uint32_t lastUsed; int32_t kernelAllocationID, searchIndex = -1, index = -1; int32_t alignedSize, shiftOffset; int32_t hr = CM_SUCCESS; kernelAllocation = stateHeap->pKernelAllocation; // Search and deallocate oldest kernel (most likely this is optimal scheduling algorithm) kernelAllocation = stateHeap->pKernelAllocation; for (kernelAllocationID = 0; kernelAllocationID < state->kernelNumInGsh; kernelAllocationID++, kernelAllocation++) { // Skip unused entries // Skip kernels flagged as locked (cannot be automatically deallocated) if (kernelAllocation->dwFlags == RENDERHAL_KERNEL_ALLOCATION_FREE || kernelAllocation->dwFlags == RENDERHAL_KERNEL_ALLOCATION_LOCKED) { continue; } // Find kernel not used for the greater amount of time (measured in number of operations) // Must not unload recently allocated kernels lastUsed = (uint32_t)(stateHeap->dwAccessCounter - kernelAllocation->dwCount); if (lastUsed > oldest) { searchIndex = kernelAllocationID; oldest = lastUsed; } } // Did not found any entry for deallocation, we get into a strange case! if (searchIndex < 0) { CM_ASSERTMESSAGE("Failed to delete any slot from GSH. It is impossible."); return CM_FAILURE; } if (stateHeap->pKernelAllocation[searchIndex].cloneKernelParams.isHeadKernel && (stateHeap->pKernelAllocation[searchIndex].cloneKernelParams.referenceCount != 0)) { // ERROR, chose a head kernel for deletion but it still has clones pointing to it return CM_FAILURE; } // Free kernel entry and states associated with the kernel (if any) kernelAllocation = &stateHeap->pKernelAllocation[searchIndex]; if (HalCm_UnloadKernel(state, kernelAllocation) != CM_SUCCESS) { CM_ASSERTMESSAGE("Failed to load kernel - no space available in GSH."); return CM_FAILURE; } // Let's check if we can merge searchIndex-1, searchIndex, searchIndex+1 index = searchIndex; PRENDERHAL_KRN_ALLOCATION kAlloc0, kAlloc1, kAlloc2; kAlloc0 = (index == 0)? nullptr : &stateHeap->pKernelAllocation[index-1]; kAlloc1 = &stateHeap->pKernelAllocation[index]; // free one kAlloc2 = (index == state->cmDeviceParam.maxGshKernelEntries - 1) ? nullptr : &stateHeap->pKernelAllocation[index + 1]; if (bIsFree(kAlloc0) && bIsFree(kAlloc2)) { // merge 3 into 1 slot and bump index after stateHeap->pKernelAllocation[index-1].dwFlags = RENDERHAL_KERNEL_ALLOCATION_FREE; state->totalKernelSize[index-1] += state->totalKernelSize[index] + state->totalKernelSize[index+1]; stateHeap->pKernelAllocation[index-1].iSize = 0; // no change for stateHeap->pKernelAllocation[index-1].dwOffset // copy the rest for (int32_t i = index + 2; ikernelNumInGsh; i++) { stateHeap->pKernelAllocation[i-2] = stateHeap->pKernelAllocation[i]; state->totalKernelSize[i-2] = state->totalKernelSize[i]; } state->kernelNumInGsh -= 2; if ( index == 0 ) HalCm_UpdateCloneKernel(state, 0, CM_SHIFT_RIGHT, 2); else HalCm_UpdateCloneKernel(state, index - 1, CM_SHIFT_RIGHT, 2); } else if (bIsFree(kAlloc0)) { // merge before and current into 1 slot stateHeap->pKernelAllocation[index-1].dwFlags = RENDERHAL_KERNEL_ALLOCATION_FREE; state->totalKernelSize[index-1] += state->totalKernelSize[index]; stateHeap->pKernelAllocation[index-1].iSize = 0; // no change for stateHeap->pKernelAllocation[index-1].dwOffset for (int32_t i = index + 1; ikernelNumInGsh; i++) { stateHeap->pKernelAllocation[i-1] = stateHeap->pKernelAllocation[i]; state->totalKernelSize[i-1] = state->totalKernelSize[i]; } state->kernelNumInGsh -= 1; if ( index == 0 ) HalCm_UpdateCloneKernel(state, 0, CM_SHIFT_RIGHT, 1); else HalCm_UpdateCloneKernel(state, index - 1, CM_SHIFT_RIGHT, 1); } else if (bIsFree(kAlloc2)) { // kAlloc0 is not free, but it can be nullptr // merge after and current into 1 slot stateHeap->pKernelAllocation[index].dwFlags = RENDERHAL_KERNEL_ALLOCATION_FREE; state->totalKernelSize[index] += state->totalKernelSize[index+1]; stateHeap->pKernelAllocation[index].iSize = 0; if (kAlloc0) { // get free space starting point alignedSize = MOS_ALIGN_CEIL(kAlloc0->iSize, 64); shiftOffset = state->totalKernelSize[index-1] - alignedSize; state->totalKernelSize[index-1] -= shiftOffset; // no change for stateHeap->pKernelAllocation[index-1].iSize -= 0; state->totalKernelSize[index] += shiftOffset; stateHeap->pKernelAllocation[index].dwOffset -= shiftOffset; } for (int32_t i = index + 1; ikernelNumInGsh; i++) { stateHeap->pKernelAllocation[i] = stateHeap->pKernelAllocation[i+1]; state->totalKernelSize[i] = state->totalKernelSize[i+1]; } state->kernelNumInGsh -= 1; if ( index == 0 ) HalCm_UpdateCloneKernel(state, 0, CM_SHIFT_RIGHT, 1); else HalCm_UpdateCloneKernel(state, index - 1, CM_SHIFT_RIGHT, 1); } else { // no merge stateHeap->pKernelAllocation[index].dwFlags = RENDERHAL_KERNEL_ALLOCATION_FREE; // no change for stateHeap->pKernelAllocation[index].iTotalSize; stateHeap->pKernelAllocation[index].iSize = 0; if(kAlloc0) { // get free space starting point alignedSize = MOS_ALIGN_CEIL(kAlloc0->iSize, 64); shiftOffset = state->totalKernelSize[index-1] - alignedSize; state->totalKernelSize[index-1] -= shiftOffset; // no change for stateHeap->pKernelAllocation[index-1].iSize -= 0; state->totalKernelSize[index] += shiftOffset; stateHeap->pKernelAllocation[index].dwOffset -= shiftOffset; } // no change for stateHeap->iNumKernels; } return hr; } /*---------------------------------------------------------------------------- | Name : HalCm_LoadKernel ( Replace RenderHal_LoadKernel) \---------------------------------------------------------------------------*/ int32_t HalCm_LoadKernel( PCM_HAL_STATE state, PCM_HAL_KERNEL_PARAM kernelParam, int32_t samplerCount, PRENDERHAL_KRN_ALLOCATION &kernelAllocation) { PRENDERHAL_STATE_HEAP stateHeap; PRENDERHAL_INTERFACE renderHal; int32_t hr; PRENDERHAL_KERNEL_PARAM parameters; PMHW_KERNEL_PARAM mhwKernelParam; int32_t kernelAllocationID; // Kernel allocation ID in GSH int32_t kernelCacheID; // Kernel cache ID int32_t kernelUniqueID; // Kernel unique ID void *kernelPtr; int32_t kernelSize; int32_t searchIndex; int32_t freeSlot; bool isClonedKernel; bool hasClones; hr = CM_SUCCESS; renderHal = state->renderHal; stateHeap = (renderHal) ? renderHal->pStateHeap : nullptr; kernelAllocationID = RENDERHAL_KERNEL_LOAD_FAIL; mhwKernelParam = &(state->kernelParamsMhw); parameters = &(state->kernelParamsRenderHal.Params); // Validate parameters if (stateHeap == nullptr || stateHeap->bIshLocked == false || stateHeap->pKernelAllocation == nullptr || kernelParam->kernelBinarySize == 0 || state->kernelNumInGsh > state->cmDeviceParam.maxGshKernelEntries) { CM_ASSERTMESSAGE("Failed to load kernel - invalid parameters."); return CM_FAILURE; } isClonedKernel = kernelParam->clonedKernelParam.isClonedKernel; hasClones = kernelParam->clonedKernelParam.hasClones; parameters->Sampler_Count = samplerCount; mhwKernelParam->iKUID = static_cast( (kernelParam->kernelId >> 32) ); mhwKernelParam->iKCID = -1; mhwKernelParam->pBinary = kernelParam->kernelBinary; mhwKernelParam->iSize = kernelParam->kernelBinarySize + CM_KERNEL_BINARY_PADDING_SIZE; // Kernel parameters kernelPtr = mhwKernelParam->pBinary; kernelSize = mhwKernelParam->iSize; kernelUniqueID = mhwKernelParam->iKUID; kernelCacheID = mhwKernelParam->iKCID; // Check if kernel is already loaded; Search free allocation index searchIndex = -1; kernelAllocation = stateHeap->pKernelAllocation; for (kernelAllocationID = 0; kernelAllocationID < state->kernelNumInGsh; kernelAllocationID++, kernelAllocation++) { if (kernelAllocation->iKUID == kernelUniqueID && kernelAllocation->iKCID == kernelCacheID) { // found match and Update kernel usage hr = HalCm_TouchKernel(state, kernelAllocationID); if (hr == CM_FAILURE) { goto finish; } // Increment reference counter mhwKernelParam->bLoaded = 1; // Record kernel allocation kernelAllocation = &stateHeap->pKernelAllocation[kernelAllocationID]; goto finish; } } if (isClonedKernel || hasClones) { hr = HalCm_InsertCloneKernel(state, kernelParam, kernelAllocation); goto finish; } // here is the algorithm // 1) search for free slot which is big enough to load current kerenel // 2) if found slot, then add current kerenel // 3) if we cannot find slot, we need to delete some entry (delete oldest first), after delete oldest entry // we will loop over to step 1 until we get enough space. // The algorithm won't fail except we load 1 kernel which is larger than 2MB do { freeSlot = CmSearchFreeSlotSize(state, mhwKernelParam, false); if (freeSlot >= 0) { // found free slot which is big enough to hold kernel hr = CmAddCurrentKernelToFreeSlot(state, freeSlot, parameters, kernelParam, mhwKernelParam, CM_NO_CLONE, -1); // update GSH states stateHeap->numKernels inside add function break; } else { if (CmDeleteOldestKernel(state, mhwKernelParam) != CM_SUCCESS) { return CM_FAILURE; } } } while(1); mhwKernelParam->bLoaded = 1; // Increment reference counter kernelAllocation = &stateHeap->pKernelAllocation[freeSlot]; // Record kernel allocation finish: return hr; } //*----------------------------------------------------------------------------- //| Purpose: Loads cloned kernel entries and kernels with clones into free slot //| Return: Result of the operation //*----------------------------------------------------------------------------- int32_t HalCm_InsertCloneKernel( PCM_HAL_STATE state, PCM_HAL_KERNEL_PARAM kernelParam, PRENDERHAL_KRN_ALLOCATION &kernelAllocation) { int32_t hr = CM_SUCCESS; int32_t kernelAllocationID; // Kernel allocation ID in GSH uint32_t tag; PMOS_INTERFACE osInterface = state->osInterface; PMHW_KERNEL_PARAM mhwKernelParam = &(state->kernelParamsMhw); int32_t freeSlot = -1; PRENDERHAL_STATE_HEAP stateHeap = state->renderHal->pStateHeap; kernelAllocation = state->renderHal->pStateHeap->pKernelAllocation; for (kernelAllocationID = 0; kernelAllocationID < state->kernelNumInGsh; kernelAllocationID++, kernelAllocation++) { if (kernelAllocation->cloneKernelParams.isHeadKernel) { if ((kernelAllocation->iKUID == kernelParam->clonedKernelParam.kernelID) || // original kernel that cloned from is already loaded as head (kernelAllocation->cloneKernelParams.cloneKernelID == kernelParam->clonedKernelParam.kernelID) || // another clone from same original kernel is serving as the head (kernelAllocation->cloneKernelParams.cloneKernelID == static_cast(kernelParam->kernelId >> 32))) // clone is serving as the head and this is the original kernel { // found match, insert 64B dummy entry and set piKAID do { // Before getting a free slot, update head kernel sync tag and count so head will not be selected for deletion // then update head kernel count after inserting clone // so that clone will be selected first for deletion (this is done in CmAddCurrentKernelToFreeSlot) // update head kernel sync tag if(state->cbbEnabled) { tag = osInterface->pfnGetGpuStatusTag(osInterface, osInterface->CurrentGpuContextOrdinal); } else { tag = state->renderHal->pStateHeap->dwNextTag; } kernelAllocation->dwSync = tag; // update the head kernel count so it will not be selected for deletion kernelAllocation->dwCount = state->renderHal->pStateHeap->dwAccessCounter++; freeSlot = CmSearchFreeSlotSize(state, mhwKernelParam, true); if (freeSlot >= 0) { // found free slot hr = CmAddCurrentKernelToFreeSlot(state, freeSlot, &(state->kernelParamsRenderHal.Params), kernelParam, &(state->kernelParamsMhw), CM_CLONE_ENTRY, kernelAllocationID); goto finish; } else { if (CmDeleteOldestKernel(state, mhwKernelParam) != CM_SUCCESS) { hr = CM_FAILURE; goto finish; } } } while (1); } } } // didn't find a match, insert this kernel as the head kernel do { freeSlot = CmSearchFreeSlotSize(state, mhwKernelParam, false); if (freeSlot >= 0) { if (kernelParam->clonedKernelParam.isClonedKernel) { hr = CmAddCurrentKernelToFreeSlot(state, freeSlot, &(state->kernelParamsRenderHal.Params), kernelParam, &(state->kernelParamsMhw), CM_CLONE_AS_HEAD_KERNEL, -1); } else { hr = CmAddCurrentKernelToFreeSlot(state, freeSlot, &(state->kernelParamsRenderHal.Params), kernelParam, &(state->kernelParamsMhw), CM_HEAD_KERNEL, -1); } break; } else { if (CmDeleteOldestKernel(state, mhwKernelParam) != CM_SUCCESS) { hr = CM_FAILURE; goto finish; } } } while (1); finish: if (hr == CM_SUCCESS) { mhwKernelParam->bLoaded = 1; kernelAllocation = &stateHeap->pKernelAllocation[freeSlot]; } return hr; } //! //! \brief Get offset to sampler state //! \details Get offset to sampler state in General State Heap, //! (Cm customized version of the RenderHal function which calculates //! the sampler offset by MDF owned parameters). //! \param PCM_HAL_STATE state //! [in] Pointer to CM_HAL_STATE structure //! \param PRENDERHAL_INTERFACE renderHal //! [in] Pointer to RenderHal Interface //! \param int mediaID //! [in] Media ID associated with sampler //! \param int samplerOffset //! [in] sampler offset from the base of current kernel's sampler heap //! \param int samplerBTI //! [in] sampler BTI //! \param unsigned long *pdwSamplerOffset //! [out] optional; offset of sampler state from GSH base //! \return MOS_STATUS //! MOS_STATUS HalCm_GetSamplerOffset( PCM_HAL_STATE state, PRENDERHAL_INTERFACE renderHal, int mediaID, unsigned int samplerOffset, unsigned int samplerBTI, PMHW_SAMPLER_STATE_PARAM samplerParam, uint32_t *pdwSamplerOffset) { PRENDERHAL_MEDIA_STATE_LEGACY pCurMediaStateLegacy = (PRENDERHAL_MEDIA_STATE_LEGACY)renderHal->pStateHeap->pCurMediaState; unsigned int tmpSamplerOffset = pCurMediaStateLegacy->pDynamicState->Sampler3D.dwOffset + state->taskParam->samplerOffsetsByKernel[mediaID] + samplerOffset; if (pdwSamplerOffset != nullptr) { *pdwSamplerOffset = tmpSamplerOffset; } if (samplerParam->SamplerType == MHW_SAMPLER_TYPE_3D) { samplerParam->Unorm.IndirectStateOffset = MOS_ALIGN_CEIL( pCurMediaStateLegacy->pDynamicState->Sampler3D.dwOffset + state->taskParam->samplerIndirectOffsetsByKernel[mediaID] + samplerBTI * renderHal->pHwSizes->dwSizeSamplerIndirectState, 1 << MHW_SAMPLER_INDIRECT_SHIFT); } return MOS_STATUS_SUCCESS; } //! //! \brief Setup Interface Descriptor //! \details Set interface descriptor, (overriding RenderHal function), //! (Cm customized version of the RenderHal function which set //! dwSamplerOffset and dwSamplerCount by MDF owned parameters). //! \param PCM_HAL_STATE state //! [in] Pointer to CM_HAL_STATE structure //! \param PRENDERHAL_INTERFACE renderHal //! [in] Pointer to HW interface //! \param PRENDERHAL_MEDIA_STATE mediaState //! [in] Pointer to media state //! \param PRENDERHAL_KRN_ALLOCATION kernelAllocation //! [in] Pointer to kernel allocation //! \param PRENDERHAL_INTERFACE_DESCRIPTOR_PARAMS interfaceDescriptorParams //! [in] Pointer to interface descriptor parameters //! \param PMHW_GPGPU_WALKER_PARAMS pGpGpuWalkerParams //! [in] Pointer to gpgpu walker parameters //! \return MOS_STATUS //! MOS_STATUS HalCm_SetupInterfaceDescriptor( PCM_HAL_STATE state, PRENDERHAL_INTERFACE renderHal, PRENDERHAL_MEDIA_STATE mediaState, PRENDERHAL_KRN_ALLOCATION kernelAllocation, PRENDERHAL_INTERFACE_DESCRIPTOR_PARAMS interfaceDescriptorParams) { MOS_STATUS eStatus = MOS_STATUS_SUCCESS; MHW_ID_ENTRY_PARAMS params; PRENDERHAL_STATE_HEAP stateHeap; PRENDERHAL_DYNAMIC_STATE dynamicState; unsigned long mediaStateOffset; PRENDERHAL_MEDIA_STATE_LEGACY mediaStateLegacy = (PRENDERHAL_MEDIA_STATE_LEGACY)mediaState; //----------------------------------------- MHW_RENDERHAL_CHK_NULL(renderHal); MHW_RENDERHAL_CHK_NULL(renderHal->pMhwStateHeap); MHW_RENDERHAL_CHK_NULL(mediaStateLegacy); MHW_RENDERHAL_CHK_NULL(mediaStateLegacy->pDynamicState); MHW_RENDERHAL_CHK_NULL(interfaceDescriptorParams); //----------------------------------------- // Get states, params stateHeap = renderHal->pStateHeap; dynamicState = mediaStateLegacy->pDynamicState; mediaStateOffset = dynamicState->memoryBlock.GetOffset(); params.dwMediaIdOffset = mediaStateOffset + dynamicState->MediaID.dwOffset; params.iMediaId = interfaceDescriptorParams->iMediaID; params.dwKernelOffset = kernelAllocation->dwOffset; params.dwSamplerOffset = mediaStateOffset + dynamicState->Sampler3D.dwOffset + state->taskParam->samplerOffsetsByKernel[params.iMediaId]; params.dwSamplerCount = ( state->taskParam->samplerCountsByKernel[params.iMediaId] + 3 ) / 4; params.dwSamplerCount = (params.dwSamplerCount > 4) ? 4 : params.dwSamplerCount; params.dwBindingTableOffset = interfaceDescriptorParams->iBindingTableID * stateHeap->iBindingTableSize; params.iCurbeOffset = interfaceDescriptorParams->iCurbeOffset; params.iCurbeLength = interfaceDescriptorParams->iCurbeLength; params.bBarrierEnable = interfaceDescriptorParams->blBarrierEnable; params.bGlobalBarrierEnable = interfaceDescriptorParams->blGlobalBarrierEnable; //It's only applied for BDW+ params.dwNumberofThreadsInGPGPUGroup = interfaceDescriptorParams->iNumberThreadsInGroup; params.dwSharedLocalMemorySize = renderHal->pfnEncodeSLMSize(renderHal, interfaceDescriptorParams->iSLMSize); params.iCrsThdConDataRdLn = interfaceDescriptorParams->iCrsThrdConstDataLn; params.memoryBlock = &dynamicState->memoryBlock; MHW_RENDERHAL_CHK_STATUS(renderHal->pMhwStateHeap->AddInterfaceDescriptorData(¶ms)); dynamicState->MediaID.iCurrent++; finish: return eStatus; } /*---------------------------------------------------------------------------- | Name : HalCm_AllocateMediaID replace old RenderHal_AllocateMediaID | Don't need touch kernel since we handle this a loadKernel time | | Purpose : Allocates an setup Interface Descriptor for Media Pipeline | | Arguments : [in] renderHal - Pointer to RenderHal interface structure | [in] kernelParam - Pointer to Kernel parameters | [in] pKernelAllocationID - Pointer to Kernel allocation | [in] bindingTableID - Binding table ID | [in] curbeOffset - Curbe offset (from CURBE base) | | Returns : Media Interface descriptor ID | -1 if invalid parameters | no Interface Descriptor entry available in GSH | | Comments : Kernel must be preloaded | Curbe must be allocated using pfnAllocateCurbe | Binding Table must be allocated using pfnAllocateBindingTable \---------------------------------------------------------------------------*/ //! //! \brief //! \details //! \param PRENDERHAL_INTERFACE renderHal //| \param PCM_HAL_KERNEL_PARAM kernelParam //| \param PRENDERHAL_KRN_ALLOCATION kernelAllocation //| \param int32_t bindingTableID //| \param int32_t curbeOffset //! \return int32_t //! int32_t HalCm_AllocateMediaID( PCM_HAL_STATE state, PCM_HAL_KERNEL_PARAM kernelParam, PRENDERHAL_KRN_ALLOCATION kernelAllocation, int32_t bindingTableID, int32_t curbeOffset) { PRENDERHAL_INTERFACE renderHal = state->renderHal; PRENDERHAL_MEDIA_STATE_LEGACY curMediaState; int32_t curbeSize, iCurbeCurrent; int32_t interfaceDescriptor; RENDERHAL_INTERFACE_DESCRIPTOR_PARAMS interfaceDescriptorParams; interfaceDescriptor = -1; // Obtain pointer and validate current media state curMediaState = (PRENDERHAL_MEDIA_STATE_LEGACY)renderHal->pStateHeap->pCurMediaState; if (state->dshEnabled) { if (curMediaState == nullptr || (state->dshEnabled && (curMediaState->pDynamicState == nullptr))) { CM_ASSERTMESSAGE("Invalid Media State."); goto finish; } } else { if (curMediaState == nullptr) { CM_ASSERTMESSAGE("Invalid Media State."); goto finish; } } // Validate kernel allocation (kernel must be pre-loaded into GSH) if (!kernelAllocation || kernelAllocation->dwFlags == RENDERHAL_KERNEL_ALLOCATION_FREE || kernelAllocation->iSize == 0) { CM_ASSERTMESSAGE("Error: Invalid Kernel Allocation."); goto finish; } // Check Curbe allocation (CURBE_Lenght is in 256-bit count -> convert to bytes) curbeSize = kernelParam->curbeSizePerThread; if (state->dshEnabled) { iCurbeCurrent = curMediaState->pDynamicState->Curbe.iCurrent; } else { iCurbeCurrent = curMediaState->iCurbeOffset; } if (curbeSize <= 0) { // Curbe is not used by the kernel curbeSize = curbeOffset = 0; } // Validate Curbe Offset (curbe must be pre-allocated) else if ( curbeOffset < 0 || // Not allocated (curbeOffset & 0x1F) != 0 || // Invalid alignment (curbeOffset + curbeSize) > iCurbeCurrent) // Invalid size { CM_ASSERTMESSAGE("Error: Invalid Curbe Allocation."); goto finish; } // Try to reuse interface descriptor (for 2nd level buffer optimizations) // Check if ID already in use by another kernel - must use a different ID interfaceDescriptor = renderHal->pfnGetMediaID(renderHal, curMediaState, kernelAllocation); if (interfaceDescriptor < 0) { CM_ASSERTMESSAGE("Error: No Interface Descriptor available."); goto finish; } interfaceDescriptorParams.iMediaID = interfaceDescriptor; interfaceDescriptorParams.iBindingTableID = bindingTableID; //CURBE size and offset setting //Media w/o group: only per-thread CURBE is used, CrossThread CURBE is not used. //Media w/ group: should follow GPGPU walker setting, there is per-thread CURBE and cross-thread CURBE. But per-thread CURBE should be ZERO, and all should be cross-thread CURBE //GPGPU: both per-thread CURBE and cross-thread CURBE need be set. interfaceDescriptorParams.iCurbeOffset = curbeOffset; if ((!kernelParam->gpgpuWalkerParams.gpgpuEnabled) && (kernelParam->kernelThreadSpaceParam.groupSelect == CM_MW_GROUP_NONE) && (state->taskParam->mediaWalkerGroupSelect == CM_MW_GROUP_NONE)) { //Media pipe without group interfaceDescriptorParams.iCurbeLength = kernelParam->curbeSizePerThread; interfaceDescriptorParams.iCrsThrdConstDataLn = kernelParam->crossThreadConstDataLen; //should always be 0 in this case interfaceDescriptorParams.iNumberThreadsInGroup = (kernelParam->numberThreadsInGroup > 0) ? kernelParam->numberThreadsInGroup : 1; // This field should not be set to 0 even if the barrier is disabled, since an accurate value is needed for proper pre-emption. interfaceDescriptorParams.blGlobalBarrierEnable = false; interfaceDescriptorParams.blBarrierEnable = false; interfaceDescriptorParams.iSLMSize = 0; } else if ((!kernelParam->gpgpuWalkerParams.gpgpuEnabled) && ((kernelParam->kernelThreadSpaceParam.groupSelect != CM_MW_GROUP_NONE) || (state->taskParam->mediaWalkerGroupSelect != CM_MW_GROUP_NONE))) { //Media w/ group interfaceDescriptorParams.iCurbeLength = 0; //No using per-thread CURBE interfaceDescriptorParams.iCrsThrdConstDataLn = kernelParam->curbeSizePerThread; //treat all CURBE as cross-thread CURBE interfaceDescriptorParams.iNumberThreadsInGroup = (kernelParam->numberThreadsInGroup > 0) ? kernelParam->numberThreadsInGroup : 1; // This field should not be set to 0 even if the barrier is disabled, since an accurate value is needed for proper pre-emption. interfaceDescriptorParams.blBarrierEnable = (kernelParam->barrierMode != CM_NO_BARRIER) ? true : false; interfaceDescriptorParams.blGlobalBarrierEnable = (kernelParam->barrierMode == CM_GLOBAL_BARRIER) ? true : false; interfaceDescriptorParams.iSLMSize = kernelParam->slmSize; } else { //GPGPU pipe interfaceDescriptorParams.iCurbeLength = kernelParam->curbeSizePerThread; interfaceDescriptorParams.iCrsThrdConstDataLn = kernelParam->crossThreadConstDataLen; interfaceDescriptorParams.iNumberThreadsInGroup = (kernelParam->numberThreadsInGroup > 0) ? kernelParam->numberThreadsInGroup : 1; interfaceDescriptorParams.blBarrierEnable = (kernelParam->barrierMode != CM_NO_BARRIER) ? true : false; interfaceDescriptorParams.blGlobalBarrierEnable = (kernelParam->barrierMode == CM_GLOBAL_BARRIER) ? true : false; interfaceDescriptorParams.iSLMSize = kernelParam->slmSize; } if (state->useNewSamplerHeap == true) { HalCm_SetupInterfaceDescriptor(state, renderHal, curMediaState, kernelAllocation, &interfaceDescriptorParams); } else { // Setup Media ID entry - this call could be HW dependent renderHal->pfnSetupInterfaceDescriptor( renderHal, curMediaState, kernelAllocation, &interfaceDescriptorParams); } finish: return interfaceDescriptor; } bool isRenderTarget(PCM_HAL_STATE state, uint32_t index) { bool readSync = false; readSync = state->umdSurf2DTable[index].readSyncs[state->osInterface->CurrentGpuContextOrdinal]; if (readSync) return false; else return true; } int32_t HalCm_DSH_LoadKernelArray( PCM_HAL_STATE state, PCM_HAL_KERNEL_PARAM *kernelArray, int32_t kernelCount, PRENDERHAL_KRN_ALLOCATION *krnAllocation) { PRENDERHAL_INTERFACE renderHal; PCM_HAL_KERNEL_PARAM kernel; PMHW_STATE_HEAP_MEMORY_BLOCK memoryBlock; // Kernel memory block int32_t totalSize; // Total size uint32_t blockSize[CM_MAX_KERNELS_PER_TASK]; // Size of kernels to load int32_t blockCount; // Number of kernels to load MOS_STATUS eStatus = MOS_STATUS_SUCCESS; int32_t hr = CM_FAILURE; renderHal = state->renderHal; state->criticalSectionDSH->Acquire(); do { blockCount = 0; totalSize = 0; // Obtain list of kernels already loaded, discard kernels loaded in older heaps. // Calculate total size of kernels to be loaded, and get size of largest kernel. for (int i = 0; i < kernelCount; i++) { // Find out if kernel is already allocated and loaded in ISH kernel = kernelArray[i]; krnAllocation[i] = (PRENDERHAL_KRN_ALLOCATION)renderHal->pfnSearchDynamicKernel(renderHal, static_cast((kernel->kernelId >> 32)), -1); // Kernel is allocated - check if kernel is in current ISH if (krnAllocation[i]) { // Check if kernel is loaded memoryBlock = krnAllocation[i]->pMemoryBlock; if (memoryBlock) { // Kernel needs to be reloaded in current heap if (memoryBlock->pStateHeap != renderHal->pMhwStateHeap->GetISHPointer() || state->forceKernelReload) //pInstructionStateHeaps { renderHal->pMhwStateHeap->FreeDynamicBlockDyn(MHW_ISH_TYPE, memoryBlock); krnAllocation[i]->pMemoryBlock = nullptr; } else { // Increment kernel usage count, used in kernel caching architecture state->dshKernelCacheHit++; krnAllocation[i]->dwCount++; // Lock kernel to avoid removal while loading other kernels krnAllocation[i]->dwFlags = RENDERHAL_KERNEL_ALLOCATION_LOCKED; } } else if (krnAllocation[i]->dwFlags == RENDERHAL_KERNEL_ALLOCATION_REMOVED) { // This is a kernel that was unloaded and now needs to be reloaded // Track how many times this "cache miss" happens to determine if the // ISH is under pressure and needs to be expanded state->dshKernelCacheMiss++; } } else { // Assign kernel allocation for this kernel krnAllocation[i] = renderHal->pfnAllocateDynamicKernel(renderHal, static_cast((kernel->kernelId >> 32)), -1); CM_CHK_NULL_GOTOFINISH_MOSERROR(krnAllocation[i]); } // Kernel is not loaded -> add to list of kernels to be loaded if (krnAllocation[i]->pMemoryBlock == nullptr && krnAllocation[i]->dwFlags != RENDERHAL_KERNEL_ALLOCATION_LOADING) { // Increment amount of data that needs to be loaded in ISH (kernel already registered but unloaded) blockSize[blockCount++] = kernel->kernelBinarySize + CM_KERNEL_BINARY_PADDING_SIZE; totalSize += kernel->kernelBinarySize + CM_KERNEL_BINARY_PADDING_SIZE; // Flag this kernel as loading - one single kernel instance is needed, not multiple! // If the same kernel is used multiple times, avoid multiple reservations/loads krnAllocation[i]->dwFlags = RENDERHAL_KERNEL_ALLOCATION_LOADING; } } // Use Hit/Miss ratio to ignore eventual cache misses // This code prevents ISH reallocation in case of eventual cache misses while (state->dshKernelCacheHit >= HAL_CM_KERNEL_CACHE_HIT_TO_MISS_RATIO) { if (state->dshKernelCacheMiss > 0) state->dshKernelCacheMiss--; state->dshKernelCacheHit -= HAL_CM_KERNEL_CACHE_HIT_TO_MISS_RATIO; } // Grow the kernel heap if too many kernels are being reloaded or there isn't enough room to load all kernels if (state->dshKernelCacheMiss > HAL_CM_KERNEL_CACHE_MISS_THRESHOLD || renderHal->pfnRefreshDynamicKernels(renderHal, totalSize, blockSize, blockCount) != MOS_STATUS_SUCCESS) { renderHal->pfnExpandKernelStateHeap(renderHal, (uint32_t)totalSize); state->dshKernelCacheHit = 0; state->dshKernelCacheMiss = 0; continue; } // blockSize/blockCount define a list of blocks that must be loaded in current ISH for the // kernels not yet present. Pre-existing kernels are marked as bStatic to avoid being unloaded here if (blockCount > 0) { // Allocate array of kernels MHW_STATE_HEAP_DYNAMIC_ALLOC_PARAMS params; params.piSizes = (int32_t*)blockSize; params.iCount = blockCount; params.dwAlignment = RENDERHAL_KERNEL_BLOCK_ALIGN; params.bHeapAffinity = true; // heap affinity - load all kernels in the same heap params.pHeapAffinity = renderHal->pMhwStateHeap->GetISHPointer(); // Select the active instruction heap params.dwScratchSpace = 0; params.bZeroAssignedMem = true; params.bStatic = true; params.bGrow = false; // Try to allocate array of blocks; if it fails, we may need to clear some space or grow the heap! memoryBlock = renderHal->pMhwStateHeap->AllocateDynamicBlockDyn(MHW_ISH_TYPE, ¶ms); if (!memoryBlock) { // Reset flags for (int i = 0; i < kernelCount; i++) { if (krnAllocation[i] && krnAllocation[i]->dwFlags == RENDERHAL_KERNEL_ALLOCATION_LOADING) { krnAllocation[i]->dwFlags = RENDERHAL_KERNEL_ALLOCATION_STALE; } } if (renderHal->pfnRefreshDynamicKernels(renderHal, totalSize, blockSize, blockCount) != MOS_STATUS_SUCCESS) { renderHal->pfnExpandKernelStateHeap(renderHal, (uint32_t)totalSize); } continue; } // All blocks are allocated in ISH // Setup kernel allocations, load kernel binaries for (int32_t i = 0; i < kernelCount; i++) { // Load kernels in ISH if (!krnAllocation[i]->pMemoryBlock) { PCM_HAL_KERNEL_PARAM kernelParam = kernelArray[i]; PRENDERHAL_KRN_ALLOCATION allocation = krnAllocation[i]; if (memoryBlock) { allocation->iKID = -1; allocation->iKUID = static_cast((kernelArray[i]->kernelId >> 32)); allocation->iKCID = -1; FrameTrackerTokenFlat_SetProducer(&allocation->trackerToken, &renderHal->trackerProducer); FrameTrackerTokenFlat_Merge(&allocation->trackerToken, renderHal->currentTrackerIndex, renderHal->trackerProducer.GetNextTracker(renderHal->currentTrackerIndex)); allocation->dwOffset = memoryBlock->dwDataOffset; allocation->iSize = kernelArray[i]->kernelBinarySize + CM_KERNEL_BINARY_PADDING_SIZE; allocation->dwCount = 0; allocation->dwFlags = RENDERHAL_KERNEL_ALLOCATION_USED; allocation->Params = state->kernelParamsRenderHal.Params; allocation->pMhwKernelParam = &state->kernelParamsMhw; allocation->pMemoryBlock = memoryBlock; // Copy kernel data // Copy MovInstruction First if (allocation->pMemoryBlock && allocation->pMemoryBlock->dwDataSize >= kernelParam->kernelBinarySize) { MOS_SecureMemcpy(allocation->pMemoryBlock->pDataPtr, kernelParam->movInsDataSize, kernelParam->movInsData, kernelParam->movInsDataSize); // Copy Cm Kernel Binary MOS_SecureMemcpy(allocation->pMemoryBlock->pDataPtr + kernelParam->movInsDataSize, kernelParam->kernelBinarySize - kernelParam->movInsDataSize, kernelParam->kernelBinary, kernelParam->kernelBinarySize - kernelParam->movInsDataSize); // Padding bytes dummy instructions after kernel binary to resolve page fault issue MOS_ZeroMemory(allocation->pMemoryBlock->pDataPtr + kernelParam->kernelBinarySize, CM_KERNEL_BINARY_PADDING_SIZE); } // Get next memory block returned as part of the array memoryBlock = memoryBlock->pNext; } } } } // Kernel load was successfull, or nothing else to load - // Quit the kernel load loop hr = CM_SUCCESS; eStatus = MOS_STATUS_SUCCESS; break; } while (1); finish: if (eStatus == MOS_STATUS_SUCCESS) { for (int32_t i = 0; i < kernelCount; i++) { renderHal->pfnTouchDynamicKernel(renderHal, krnAllocation[i]); } } state->criticalSectionDSH->Release(); return hr; } MOS_STATUS HalCm_DSH_GetDynamicStateConfiguration( PCM_HAL_STATE state, PRENDERHAL_DYNAMIC_MEDIA_STATE_PARAMS params, uint32_t numKernels, PCM_HAL_KERNEL_PARAM *kernels, uint32_t *piCurbeOffsets) { PCM_HAL_KERNEL_PARAM cmKernel; PRENDERHAL_INTERFACE renderHal = state->renderHal; PRENDERHAL_KRN_ALLOCATION krnAllocation; MOS_ZeroMemory(params, sizeof(RENDERHAL_DYNAMIC_MEDIA_STATE_PARAMS)); params->iMaxMediaIDs = numKernels; for (uint32_t i = 0; i < numKernels; i++) { cmKernel = kernels[i]; // get max curbe size int32_t curbeSize = MOS_ALIGN_CEIL(cmKernel->totalCurbeSize, state->renderHal->dwCurbeBlockAlign); int32_t curbeOffset = piCurbeOffsets[i] + curbeSize; params->iMaxCurbeOffset = MOS_MAX(params->iMaxCurbeOffset, curbeOffset); params->iMaxCurbeSize += curbeSize; // get max spill size params->iMaxSpillSize = MOS_MAX(params->iMaxSpillSize, (int32_t)cmKernel->spillSize); // check if kernel already used - increase Max Media ID to allow BB reuse logic krnAllocation = renderHal->pfnSearchDynamicKernel(renderHal, static_cast((cmKernel->kernelId >> 32)), -1); if (krnAllocation) { params->iMaxMediaIDs = MOS_MAX(params->iMaxMediaIDs, krnAllocation->iKID + 1); } } if (state->useNewSamplerHeap == true) { // Update offset to the base of first kernel and update count // for 3D sampler, update indirect state information unsigned int heapOffset = 0; unsigned int sampler3DCount = 0; MHW_SAMPLER_STATE_PARAM samplerParamMhw = {}; SamplerParam samplerParam = {}; samplerParamMhw.SamplerType = MHW_SAMPLER_TYPE_3D; state->cmHalInterface->GetSamplerParamInfoForSamplerType(&samplerParamMhw, samplerParam); for (unsigned int i = 0; i < numKernels; i++) { cmKernel = kernels[i]; std::list *sampler_heap = cmKernel->samplerHeap; std::list::iterator iter; heapOffset = MOS_ALIGN_CEIL(heapOffset, MHW_SAMPLER_STATE_ALIGN); state->taskParam->samplerOffsetsByKernel[i] = heapOffset; state->taskParam->samplerCountsByKernel[i] = sampler_heap->size(); if (sampler_heap->size() > 0) { heapOffset = heapOffset + sampler_heap->back().heapOffset + sampler_heap->back().size; // 3D sampler needs indirect sampler heap, so calculates the required size // and offset for indirect sampler heap. unsigned int max3DCount = 0; for (iter = sampler_heap->begin(); iter != sampler_heap->end(); ++iter) { if (iter->elementType == samplerParam.elementType) { if (iter->userDefinedBti == true) { max3DCount = iter->bti + 1; } else { max3DCount += 1; } } } heapOffset = MOS_ALIGN_CEIL(heapOffset, MHW_SAMPLER_STATE_ALIGN); state->taskParam->samplerIndirectOffsetsByKernel[i] = heapOffset; heapOffset += max3DCount * state->renderHal->pHwSizes->dwSizeSamplerIndirectState; sampler3DCount += max3DCount; } } // Temporary solution for DSH sampler heap assginment: // Adjust sampler space for DSH, because the DSH use sampler count to // allocate the space. However the mechanism is not correct. The sampler // heap size is actually calculated by the maximum offset of the largest // sampler type. // So the offset of largest element plus the size of all of the largest // element samplers should be equal to the maximum size. However we cannot // do this because of the DSH's mechanism. // To resolve this, we first let DSH allocate enough 3D samplers // (because 3D samplers has indirect state), then just convert the rest of // the heap to AVS. Here we only care about the size, not the correct // number because we are going to calculate the offset by ourself. // Since DSH allocation has some alignments inside, the actually size of the // heap should be slightly larger, which should be OK. samplerParamMhw.SamplerType = MHW_SAMPLER_TYPE_AVS; state->cmHalInterface->GetSamplerParamInfoForSamplerType(&samplerParamMhw, samplerParam); params->iMaxSamplerIndex3D = (sampler3DCount + numKernels - 1) / numKernels; params->iMaxSamplerIndexAVS = ((heapOffset - sampler3DCount * (state->renderHal->pHwSizes->dwSizeSamplerState + state->renderHal->pHwSizes->dwSizeSamplerIndirectState)) + samplerParam.btiMultiplier * numKernels - 1) / (samplerParam.btiMultiplier * numKernels); } else { // Get total sampler count // Initialize pointers to samplers and reset sampler index table MOS_FillMemory(state->samplerIndexTable, state->cmDeviceParam.maxSamplerTableSize, CM_INVALID_INDEX); params->iMaxSamplerIndex3D = CM_MAX_3D_SAMPLER_SIZE; params->iMaxSamplerIndexAVS = CM_MAX_AVS_SAMPLER_SIZE; params->iMaxSamplerIndexConv = 0; params->iMaxSamplerIndexMisc = 0; params->iMax8x8Tables = CM_MAX_AVS_SAMPLER_SIZE; } return MOS_STATUS_SUCCESS; } MOS_STATUS HalCm_DSH_UnregisterKernel( PCM_HAL_STATE state, uint64_t kernelId) { PRENDERHAL_INTERFACE renderHal = state->renderHal; PRENDERHAL_KRN_ALLOCATION krnAllocation = renderHal->pfnSearchDynamicKernel(renderHal, static_cast((kernelId >> 32)), -1); if (krnAllocation) { state->criticalSectionDSH->Acquire(); renderHal->pfnUnregisterKernel(renderHal, krnAllocation); state->criticalSectionDSH->Release(); } return MOS_STATUS_SUCCESS; } //*----------------------------------------------------------------------------- //| Purpose: Setup Sampler State //| Returns: Result of the operation //*----------------------------------------------------------------------------- MOS_STATUS HalCm_SetupSamplerState( PCM_HAL_STATE state, PCM_HAL_KERNEL_PARAM kernelParam, PCM_HAL_KERNEL_ARG_PARAM argParam, PCM_HAL_INDEX_PARAM indexParam, int32_t mediaID, uint32_t threadIndex, uint8_t *buffer) { MOS_STATUS eStatus; PRENDERHAL_INTERFACE renderHal; PMHW_SAMPLER_STATE_PARAM samplerParam; uint8_t *src; uint8_t *dst; uint32_t index; uint32_t samplerIndex = 0; void *sampler = nullptr; uint32_t samplerOffset = 0; eStatus = MOS_STATUS_SUCCESS; CM_CHK_NULL_GOTOFINISH_MOSERROR(state); renderHal = state->renderHal; if (indexParam->samplerIndexCount >= (uint32_t)renderHal->StateHeapSettings.iSamplers) { eStatus = MOS_STATUS_INVALID_PARAMETER; CM_ASSERTMESSAGE( "Exceeded Max samplers '%d'", indexParam->samplerIndexCount); goto finish; } // Get the Index to sampler array from the kernel data //---------------------------------- CM_ASSERT(argParam->unitSize == sizeof(index)); //---------------------------------- src = argParam->firstValue + (threadIndex * argParam->unitSize); index = *((uint32_t*)src); // check to see if the data present for the sampler in the array if (index >= state->cmDeviceParam.maxSamplerTableSize || !state->samplerTable[index].bInUse) { eStatus = MOS_STATUS_INVALID_PARAMETER; CM_ASSERTMESSAGE( "Invalid Sampler array index '%d'", index); goto finish; } // Setup samplers samplerParam = &state->samplerTable[index]; if (state->useNewSamplerHeap == true) { std::list::iterator iter; for (iter = kernelParam->samplerHeap->begin(); iter != kernelParam->samplerHeap->end(); ++iter) { if ((iter->samplerTableIndex == index)&&(iter->regularBti == true)) { break; } } if (iter != kernelParam->samplerHeap->end()) { samplerIndex = iter->bti; } else { // There must be incorrect internal logic CM_ASSERTMESSAGE( "BTI calculation error in cm_hal\n"); return MOS_STATUS_UNKNOWN; } HalCm_GetSamplerOffset(state, renderHal, mediaID, iter->heapOffset, iter->bti, samplerParam, &samplerOffset); } else { // Check to see if sampler is already assigned samplerIndex = state->samplerIndexTable[index]; if ((int)samplerIndex == CM_INVALID_INDEX) { switch (state->samplerTable[index].ElementType) { case MHW_Sampler2Elements: { unsigned int index = 0; index = state->samplerStatistics.samplerIndexBase[MHW_Sampler2Elements]; while (state->samplerIndexTable[index] != CM_INVALID_INDEX) { index++; } samplerIndex = index; state->samplerStatistics.samplerIndexBase[MHW_Sampler2Elements] = (index + 1); break; } case MHW_Sampler4Elements: { unsigned int index = 0; index = state->samplerStatistics.samplerIndexBase[MHW_Sampler4Elements]; while (state->samplerIndexTable[index] != CM_INVALID_INDEX) { index++; } samplerIndex = index; state->samplerStatistics.samplerIndexBase[MHW_Sampler4Elements] = (index + 1); break; } case MHW_Sampler8Elements: { unsigned int index = 0; index = state->samplerStatistics.samplerIndexBase[MHW_Sampler8Elements]; while (state->samplerIndexTable[index] != CM_INVALID_INDEX) { index++; } samplerIndex = index; state->samplerStatistics.samplerIndexBase[MHW_Sampler8Elements] = (index + 1); break; } case MHW_Sampler64Elements: { unsigned int index = 0; index = state->samplerStatistics.samplerIndexBase[MHW_Sampler64Elements]; while (state->samplerIndexTable[index] != CM_INVALID_INDEX) { index += index + 2; } samplerIndex = index; state->samplerStatistics.samplerIndexBase[MHW_Sampler64Elements] = (index + 2); break; } case MHW_Sampler128Elements: { unsigned int index = 0; index = state->samplerStatistics.samplerIndexBase[MHW_Sampler128Elements]; while (state->samplerIndexTable[index] != CM_INVALID_INDEX) { index++; } samplerIndex = index; state->samplerStatistics.samplerIndexBase[MHW_Sampler128Elements] = (index + 1); break; } default: CM_ASSERTMESSAGE("Invalid sampler type '%d'.", state->samplerTable[index].SamplerType); break; } } CM_CHK_MOSSTATUS_GOTOFINISH(renderHal->pfnGetSamplerOffsetAndPtr( renderHal, mediaID, samplerIndex, samplerParam, &samplerOffset, &sampler)); } CM_CHK_MOSSTATUS_GOTOFINISH(renderHal->pMhwStateHeap->AddSamplerStateData( samplerOffset, &(((PRENDERHAL_MEDIA_STATE_LEGACY)renderHal->pStateHeap->pCurMediaState)->pDynamicState->memoryBlock), samplerParam)); state->samplerIndexTable[index] = (unsigned char)samplerIndex; // Update the Batch Buffer if (buffer) { dst = buffer + argParam->payloadOffset; *((uint32_t*)dst) = samplerIndex; } finish: return eStatus; } //*----------------------------------------------------------------------------- //| Purpose: Setup Sampler State //| Returns: Result of the operation //*----------------------------------------------------------------------------- MOS_STATUS HalCm_SetupSamplerStateWithBTIndex( PCM_HAL_STATE state, PCM_HAL_KERNEL_PARAM kernelParam, PCM_HAL_SAMPLER_BTI_ENTRY samplerBTIEntry, uint32_t samplerCount, int32_t mediaID ) { MOS_STATUS eStatus = MOS_STATUS_SUCCESS; PRENDERHAL_INTERFACE renderHal; PMHW_SAMPLER_STATE_PARAM samplerParam; uint32_t index; uint32_t samplerIndex; void *sampler = nullptr; uint32_t samplerOffset = 0; renderHal = state->renderHal; if (state->useNewSamplerHeap != true) { if (samplerCount >= (uint32_t)renderHal->StateHeapSettings.iSamplers) { eStatus = MOS_STATUS_INVALID_PARAMETER; CM_ASSERTMESSAGE( "Exceeded Max samplers '%d'", samplerCount); goto finish; } } index = samplerBTIEntry[ samplerCount ].samplerIndex; // check to see if the data present for the sampler in the array if ( index >= state->cmDeviceParam.maxSamplerTableSize || !state->samplerTable[ index ].bInUse ) { eStatus = MOS_STATUS_INVALID_PARAMETER; CM_ASSERTMESSAGE( "Invalid Sampler array index '%d'", index ); goto finish; } samplerIndex = samplerBTIEntry[ samplerCount ].samplerBTI; // Setup samplers samplerParam = &state->samplerTable[ index ]; if (state->useNewSamplerHeap == true) { std::list::iterator iter; for (iter = kernelParam->samplerHeap->begin(); iter != kernelParam->samplerHeap->end(); ++iter) { if ((iter->samplerTableIndex == index) && (iter->bti == samplerIndex) && (iter->userDefinedBti == true)) { break; } } if (iter == kernelParam->samplerHeap->end()) { // There must be incorrect internal logic CM_ASSERTMESSAGE("BTI calculation error in cm_hal\n"); return MOS_STATUS_UNKNOWN; } HalCm_GetSamplerOffset(state, renderHal, mediaID, iter->heapOffset, iter->bti, samplerParam, &samplerOffset); } else { CM_CHK_MOSSTATUS_GOTOFINISH(renderHal->pfnGetSamplerOffsetAndPtr(renderHal, mediaID, samplerIndex, samplerParam, &samplerOffset, &sampler)); } CM_CHK_MOSSTATUS_GOTOFINISH(renderHal->pMhwStateHeap->AddSamplerStateData( samplerOffset, &(((PRENDERHAL_MEDIA_STATE_LEGACY)renderHal->pStateHeap->pCurMediaState)->pDynamicState->memoryBlock), samplerParam)); finish: return eStatus; } //*----------------------------------------------------------------------------- //| Purpose: Setup Buffer surface State //| Returns: Result of the operation //*----------------------------------------------------------------------------- MOS_STATUS HalCm_SetupBufferSurfaceState( PCM_HAL_STATE state, PCM_HAL_KERNEL_ARG_PARAM argParam, PCM_HAL_INDEX_PARAM indexParam, int32_t bindingTable, int16_t globalSurface, uint32_t threadIndex, uint8_t *buffer) { MOS_STATUS eStatus; RENDERHAL_SURFACE surface; PMOS_SURFACE mosSurface; RENDERHAL_SURFACE_STATE_PARAMS surfaceParam; PRENDERHAL_INTERFACE renderHal; PRENDERHAL_SURFACE_STATE_ENTRY surfaceEntry; uint8_t *src; uint8_t *dst; uint32_t index; uint32_t btIndex; uint16_t memObjCtl; uint32_t offsetSrc; PRENDERHAL_STATE_HEAP stateHeap; CM_SURFACE_BTI_INFO surfBTIInfo; eStatus = MOS_STATUS_UNKNOWN; renderHal = state->renderHal; //GT-PIN PCM_HAL_TASK_PARAM taskParam = state->taskParam; // Get the Index to Buffer array from the kernel data CM_ASSERT(argParam->unitSize == sizeof(index)); //Init surfBTIInfo state->cmHalInterface->GetHwSurfaceBTIInfo(&surfBTIInfo); src = argParam->firstValue + (threadIndex * argParam->unitSize); index = *((uint32_t*)src) & CM_SURFACE_MASK; if (index == CM_NULL_SURFACE) { if (buffer) { dst = buffer + argParam->payloadOffset; *((uint32_t*)dst) = CM_NULL_SURFACE_BINDING_INDEX; } eStatus = MOS_STATUS_SUCCESS; goto finish; } memObjCtl = state->bufferTable[index].memObjCtl; if (!memObjCtl) { memObjCtl = CM_DEFAULT_CACHE_TYPE; } // check to see if index is valid if (index >= state->cmDeviceParam.maxBufferTableSize || (state->bufferTable[index].size == 0)) { eStatus = MOS_STATUS_INVALID_PARAMETER; CM_ASSERTMESSAGE( "Invalid Buffer surface array index '%d'", index); goto finish; } // Check to see if buffer is already assigned btIndex = state->btiBufferIndexTable[index].BTI.regularSurfIndex; if (btIndex == ( unsigned char )CM_INVALID_INDEX || argParam->aliasCreated == true) { if (globalSurface < 0) { btIndex = HalCm_GetFreeBindingIndex(state, indexParam, 1); } else { btIndex = globalSurface + surfBTIInfo.reservedSurfaceStart; //CM_BINDING_START_INDEX_OF_GLOBAL_SURFACE(state); if ( btIndex >= (surfBTIInfo.reservedSurfaceStart + CM_MAX_GLOBAL_SURFACE_NUMBER) ) { eStatus = MOS_STATUS_INVALID_PARAMETER; CM_ASSERTMESSAGE("Exceeded Max Global Surfaces '%d'", btIndex); goto finish; } } // Get Details of Buffer surface and fill the surface CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_GetSurfaceAndRegister(state, &surface, CM_ARGUMENT_SURFACEBUFFER, index, 0)); MOS_ZeroMemory(&surfaceParam, sizeof(surfaceParam)); // override the buffer offset and size if alias is used mosSurface = &(surface.OsSurface); if (state->bufferTable[index].surfaceStateEntry[argParam->aliasIndex / state->surfaceArraySize].surfaceStateSize) { mosSurface->dwWidth = state->bufferTable[index].surfaceStateEntry[argParam->aliasIndex / state->surfaceArraySize].surfaceStateSize; mosSurface->dwOffset = state->bufferTable[index].surfaceStateEntry[argParam->aliasIndex / state->surfaceArraySize].surfaceStateOffset; surface.rcSrc.right = mosSurface->dwWidth; surface.rcDst.right = mosSurface->dwWidth; } // override the mocs value if it is set if (state->bufferTable[index].surfaceStateEntry[argParam->aliasIndex / state->surfaceArraySize].surfaceStateMOCS) { memObjCtl = state->bufferTable[index].surfaceStateEntry[argParam->aliasIndex / state->surfaceArraySize].surfaceStateMOCS; } //Cache configurations state->cmHalInterface->HwSetSurfaceMemoryObjectControl(memObjCtl, &surfaceParam); // Set the isOutput by default surfaceParam.isOutput = true; // Setup Buffer surface CM_CHK_MOSSTATUS_GOTOFINISH(renderHal->pfnSetupBufferSurfaceState( renderHal, &surface, &surfaceParam, &surfaceEntry)); // Bind the surface State CM_ASSERT(((int32_t)btIndex) < renderHal->StateHeapSettings.iSurfacesPerBT + surfBTIInfo.normalSurfaceStart); CM_CHK_MOSSTATUS_GOTOFINISH(renderHal->pfnBindSurfaceState( renderHal, bindingTable, btIndex, surfaceEntry)); if ((taskParam->surfEntryInfoArrays.kernelNum != 0) && (taskParam->surfEntryInfoArrays.surfEntryInfosArray != nullptr)) { //GT-Pin uint32_t dummy = 0; CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_GetSurfaceDetails( state, indexParam, btIndex, surface.OsSurface, globalSurface, nullptr, dummy, surfaceParam, CM_ARGUMENT_SURFACEBUFFER)); } // Update index to table state->btiBufferIndexTable[ index ].BTI.regularSurfIndex = btIndex; state->btiBufferIndexTable[ index ].nPlaneNumber = 1; stateHeap = renderHal->pStateHeap; offsetSrc = ( stateHeap->iCurSshBufferIndex * stateHeap->dwSshIntanceSize ) + // Points to the Base of Current SSH Buffer Instance ( stateHeap->iBindingTableOffset ) + // Moves the pointer to Base of Array of Binding Tables ( bindingTable * stateHeap->iBindingTableSize ) + // Moves the pointer to a Particular Binding Table ( btIndex * sizeof( uint32_t ) ); // Move the pointer to correct entry state->btiBufferIndexTable[ index ].BTITableEntry.regularBtiEntryPosition = stateHeap->pSshBuffer + offsetSrc; } else { stateHeap = renderHal->pStateHeap; // Get Offset to Current Binding Table uint32_t offsetCurrentBTStart = ( stateHeap->iCurSshBufferIndex * stateHeap->dwSshIntanceSize ) + // Points to the Base of Current SSH Buffer Instance ( stateHeap->iBindingTableOffset ) + // Moves the pointer to Base of Array of Binding Tables ( bindingTable * stateHeap->iBindingTableSize ); // Moves the pointer to a Particular Binding Table uint32_t *currentBTStart = ( uint32_t *)( stateHeap->pSshBuffer + offsetCurrentBTStart ); int nEntryIndex = (int) ((uint32_t*)( state->btiBufferIndexTable[ index ].BTITableEntry.regularBtiEntryPosition ) - currentBTStart); if ( ( nEntryIndex < 0 ) || ( nEntryIndex >= renderHal->StateHeapSettings.iSurfacesPerBT ) ) { uint32_t surfaceEntries = state->btiBufferIndexTable[ index ].nPlaneNumber; if ( globalSurface < 0 ) { btIndex = HalCm_GetFreeBindingIndex( state, indexParam, surfaceEntries ); } else { btIndex = globalSurface + surfBTIInfo.reservedSurfaceStart; if ( btIndex >= (surfBTIInfo.reservedSurfaceStart + CM_MAX_GLOBAL_SURFACE_NUMBER ) ) { eStatus = MOS_STATUS_INVALID_PARAMETER; CM_ASSERTMESSAGE( "Exceeded Max Global Surfaces '%d'", btIndex ); goto finish; } } // Bind the surface State CM_ASSERT( ( ( int32_t )btIndex ) < renderHal->StateHeapSettings.iSurfacesPerBT + surfBTIInfo.normalSurfaceStart); // Get Offset to Current Binding Table uint32_t offsetDst = offsetCurrentBTStart + ( btIndex * sizeof( uint32_t ) ); // Move the pointer to correct entry uint32_t *bindingTableEntry = ( uint32_t *)( stateHeap->pSshBuffer + offsetDst ); MOS_SecureMemcpy( bindingTableEntry, sizeof( uint32_t ) * surfaceEntries, state->btiBufferIndexTable[ index ].BTITableEntry.regularBtiEntryPosition, sizeof( uint32_t ) * surfaceEntries ); // Update index to table state->btiBufferIndexTable[ index ].BTI.regularSurfIndex = btIndex; state->btiBufferIndexTable[ index ].BTITableEntry.regularBtiEntryPosition = bindingTableEntry; } } // Update the Batch Buffer if (buffer) { dst = buffer + argParam->payloadOffset; *((uint32_t*)dst) = btIndex; } eStatus = MOS_STATUS_SUCCESS; finish: return eStatus; } //*----------------------------------------------------------------------------- //| Purpose: Setup 3D surface State //| Returns: Result of the operation //*----------------------------------------------------------------------------- MOS_STATUS HalCm_Setup3DSurfaceState( PCM_HAL_STATE state, PCM_HAL_KERNEL_ARG_PARAM argParam, PCM_HAL_INDEX_PARAM indexParam, int32_t bindingTable, uint32_t threadIndex, uint8_t *buffer) { MOS_STATUS eStatus; PRENDERHAL_INTERFACE renderHal; RENDERHAL_SURFACE surface; RENDERHAL_SURFACE_STATE_PARAMS surfaceParam; PRENDERHAL_SURFACE_STATE_ENTRY surfaceEntries[MHW_MAX_SURFACE_PLANES]; RENDERHAL_GET_SURFACE_INFO info; uint8_t *src; uint8_t *dst; int32_t nSurfaceEntries; uint32_t index; uint32_t btIndex; uint16_t memObjCtl; uint32_t i; uint32_t offsetSrc; PRENDERHAL_STATE_HEAP stateHeap; CM_SURFACE_BTI_INFO surfBTIInfo; eStatus = MOS_STATUS_UNKNOWN; renderHal = state->renderHal; //GT-PIN PCM_HAL_TASK_PARAM taskParam = state->taskParam; state->cmHalInterface->GetHwSurfaceBTIInfo(&surfBTIInfo); // Get the Index to 3dsurface array from the kernel data CM_ASSERT(argParam->unitSize == sizeof(index)); src = argParam->firstValue + (threadIndex * argParam->unitSize); index = *((uint32_t*)src) & CM_SURFACE_MASK; if (index == CM_NULL_SURFACE) { if (buffer) { dst = buffer + argParam->payloadOffset; *((uint32_t*)dst) = CM_NULL_SURFACE_BINDING_INDEX; } eStatus = MOS_STATUS_SUCCESS; goto finish; } memObjCtl = state->surf3DTable[index].memObjCtl; if (!memObjCtl) { memObjCtl = CM_DEFAULT_CACHE_TYPE; } // check to see if the data present for the 3d surface in the array if ((index >= state->cmDeviceParam.max3DSurfaceTableSize) || Mos_ResourceIsNull(&state->surf3DTable[index].osResource)) { eStatus = MOS_STATUS_INVALID_PARAMETER; CM_ASSERTMESSAGE( "Invalid 2D surface array index '%d'", index); goto finish; } // Check to see if surface is already assigned btIndex = state->bti3DIndexTable[index].BTI.regularSurfIndex; if ( btIndex == ( unsigned char )CM_INVALID_INDEX ) { uint32_t tempPlaneIndex = 0; nSurfaceEntries = 0; // Get Details of 3D surface and fill the surface CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_GetSurfaceAndRegister(state, &surface, CM_ARGUMENT_SURFACE3D, index, 0)); // Setup 3D surface MOS_ZeroMemory(&surfaceParam, sizeof(surfaceParam)); surfaceParam.Type = renderHal->SurfaceTypeDefault; surfaceParam.Boundary = RENDERHAL_SS_BOUNDARY_ORIGINAL; surfaceParam.isOutput = true; //Cache configurations state->cmHalInterface->HwSetSurfaceMemoryObjectControl(memObjCtl, &surfaceParam); CM_CHK_MOSSTATUS_GOTOFINISH(renderHal->pfnSetupSurfaceState( renderHal, &surface, &surfaceParam, &nSurfaceEntries, surfaceEntries, nullptr)); MOS_ZeroMemory(&info, sizeof(RENDERHAL_GET_SURFACE_INFO)); CM_CHK_MOSSTATUS_GOTOFINISH(RenderHal_GetSurfaceInfo( state->osInterface, &info, &surface.OsSurface)); btIndex = HalCm_GetFreeBindingIndex(state, indexParam, nSurfaceEntries); for (i = 0; i < (uint32_t)nSurfaceEntries; i++) { *(surfaceEntries[i]->pSurface) = surface.OsSurface; // Bind the surface State CM_CHK_MOSSTATUS_GOTOFINISH(renderHal->pfnBindSurfaceState( renderHal, bindingTable, btIndex + i, surfaceEntries[i])); if ((taskParam->surfEntryInfoArrays.kernelNum != 0) && (taskParam->surfEntryInfoArrays.surfEntryInfosArray != nullptr)) { CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_GetSurfaceDetails( state, indexParam, btIndex + i, surface.OsSurface, 0, surfaceEntries[i], tempPlaneIndex, surfaceParam, CM_ARGUMENT_SURFACE3D)); } } // Update index to table state->bti3DIndexTable[ index ].BTI.regularSurfIndex = btIndex; state->bti3DIndexTable[ index ].nPlaneNumber = nSurfaceEntries; stateHeap = renderHal->pStateHeap; offsetSrc = ( stateHeap->iCurSshBufferIndex * stateHeap->dwSshIntanceSize ) + // Points to the Base of Current SSH Buffer Instance ( stateHeap->iBindingTableOffset ) + // Moves the pointer to Base of Array of Binding Tables ( bindingTable * stateHeap->iBindingTableSize ) + // Moves the pointer to a Particular Binding Table ( btIndex * sizeof( uint32_t ) ); // Move the pointer to correct entry state->bti3DIndexTable[ index ].BTITableEntry.regularBtiEntryPosition = stateHeap->pSshBuffer + offsetSrc; } else { stateHeap = renderHal->pStateHeap; // Get Offset to Current Binding Table uint32_t offsetCurrentBTStart = ( stateHeap->iCurSshBufferIndex * stateHeap->dwSshIntanceSize ) + // Points to the Base of Current SSH Buffer Instance ( stateHeap->iBindingTableOffset ) + // Moves the pointer to Base of Array of Binding Tables ( bindingTable * stateHeap->iBindingTableSize ); // Moves the pointer to a Particular Binding Table uint32_t *currentBTStart = ( uint32_t *)( stateHeap->pSshBuffer + offsetCurrentBTStart ); int nEntryIndex = (int)((uint32_t*)( state->bti3DIndexTable[ index ].BTITableEntry.regularBtiEntryPosition ) - currentBTStart); if ( ( nEntryIndex < 0 ) || ( nEntryIndex >= renderHal->StateHeapSettings.iSurfacesPerBT ) ) { nSurfaceEntries = state->bti3DIndexTable[ index ].nPlaneNumber; btIndex = HalCm_GetFreeBindingIndex( state, indexParam, nSurfaceEntries ); // Bind the surface State CM_ASSERT( ( ( int32_t )btIndex ) < renderHal->StateHeapSettings.iSurfacesPerBT + surfBTIInfo.normalSurfaceStart); // Get Offset to Current Binding Table uint32_t offsetDst = offsetCurrentBTStart + ( btIndex * sizeof( uint32_t ) ); // Move the pointer to correct entry uint32_t *bindingTableEntry = ( uint32_t *)( stateHeap->pSshBuffer + offsetDst ); MOS_SecureMemcpy( bindingTableEntry, sizeof( uint32_t ) * nSurfaceEntries, state->bti3DIndexTable[ index ].BTITableEntry.regularBtiEntryPosition, sizeof( uint32_t ) * nSurfaceEntries ); // Update index to table state->bti3DIndexTable[ index ].BTI.regularSurfIndex = btIndex; state->bti3DIndexTable[ index ].BTITableEntry.regularBtiEntryPosition = bindingTableEntry; } } // Update the Batch Buffer if (buffer) { dst = buffer + argParam->payloadOffset; *((uint32_t*)dst) = btIndex; } eStatus = MOS_STATUS_SUCCESS; finish: return eStatus; } /*---------------------------------------------------------------------------- | Purpose : Set's surface state interlaced settings | Returns : dword value \---------------------------------------------------------------------------*/ MOS_STATUS HalCm_HwSetSurfaceProperty( PCM_HAL_STATE state, CM_FRAME_TYPE frameType, PRENDERHAL_SURFACE_STATE_PARAMS params) { MOS_STATUS eStatus = MOS_STATUS_SUCCESS; switch (frameType) { case CM_FRAME: params->bVertStride = 0; params->bVertStrideOffs = 0; break; case CM_TOP_FIELD: params->bVertStride = 1; params->bVertStrideOffs = 0; break; case CM_BOTTOM_FIELD: params->bVertStride = 1; params->bVertStrideOffs = 1; break; default: eStatus = MOS_STATUS_UNKNOWN; } return eStatus; } // A special treatment of NV12 format. Offset of the UV plane in an NV12 surface is adjusted, so // this plane can be accessed as a separate R8G8 surface in kernels. static bool UpdateSurfaceAliasPlaneOffset( CM_HAL_SURFACE2D_SURFACE_STATE_PARAM *surfaceStateParam, MOS_SURFACE *mosSurface) { if (Format_R8G8UN != surfaceStateParam->format || Format_NV12 != mosSurface->Format) { mosSurface->Format = static_cast(surfaceStateParam->format); return false; // No need to update offset. } mosSurface->dwOffset = mosSurface->UPlaneOffset.iSurfaceOffset; mosSurface->Format = Format_R8G8UN; return false; } //*----------------------------------------------------------------------------- //| Purpose: Setup 2D surface State //| Returns: Result of the operation //*----------------------------------------------------------------------------- MOS_STATUS HalCm_Setup2DSurfaceStateBasic( PCM_HAL_STATE state, PCM_HAL_KERNEL_ARG_PARAM argParam, PCM_HAL_INDEX_PARAM indexParam, int32_t bindingTable, uint32_t threadIndex, bool pixelPitch, uint8_t *buffer, bool multipleBinding ) { MOS_STATUS eStatus; RENDERHAL_SURFACE renderHalSurface; PMOS_SURFACE surface; RENDERHAL_SURFACE_STATE_PARAMS surfaceParam; PRENDERHAL_INTERFACE renderHal; PRENDERHAL_SURFACE_STATE_ENTRY surfaceEntries[ MHW_MAX_SURFACE_PLANES ]; uint8_t *src; uint8_t *dst; int32_t nSurfaceEntries = 0; uint32_t index; uint32_t btIndex; uint16_t memObjCtl; uint32_t i; uint32_t tempPlaneIndex = 0; uint32_t offsetSrc; PRENDERHAL_STATE_HEAP stateHeap; PCM_HAL_SURFACE2D_SURFACE_STATE_PARAM surfStateParam = nullptr; UNUSED(multipleBinding); eStatus = MOS_STATUS_UNKNOWN; renderHal = state->renderHal; MOS_ZeroMemory(&renderHalSurface, sizeof(renderHalSurface)); surface = &renderHalSurface.OsSurface; nSurfaceEntries = 0; //GT-PIN PCM_HAL_TASK_PARAM taskParam = state->taskParam; // Get the Index to 2dsurface array from the kernel data CM_ASSERT( argParam->unitSize == sizeof( index ) ); src = argParam->firstValue + ( threadIndex * argParam->unitSize ); index = *( ( uint32_t *)src ) & CM_SURFACE_MASK; if ( index == CM_NULL_SURFACE ) { if ( buffer ) { dst = buffer + argParam->payloadOffset; *( ( uint32_t *)dst ) = CM_NULL_SURFACE_BINDING_INDEX; } eStatus = MOS_STATUS_SUCCESS; goto finish; } memObjCtl = state->umdSurf2DTable[index].memObjCtl; if ( !memObjCtl ) { memObjCtl = CM_DEFAULT_CACHE_TYPE; } // check to see if the data present for the 2d surface in the array if ( index >= state->cmDeviceParam.max2DSurfaceTableSize || Mos_ResourceIsNull( &state->umdSurf2DTable[ index ].osResource ) ) { eStatus = MOS_STATUS_INVALID_PARAMETER; CM_ASSERTMESSAGE( "Invalid 2D surface array index '%d'", index ); goto finish; } // Check to see if surface is already assigned unsigned char nBTIRegularSurf, nBTISamplerSurf; nBTIRegularSurf = state->bti2DIndexTable[ index ].BTI.regularSurfIndex; nBTISamplerSurf = state->bti2DIndexTable[ index ].BTI.samplerSurfIndex; if (((!pixelPitch && (nBTIRegularSurf != (unsigned char)CM_INVALID_INDEX)) || (pixelPitch && (nBTISamplerSurf != (unsigned char)CM_INVALID_INDEX))) && argParam->aliasCreated == false ) { if ( pixelPitch ) { btIndex = nBTISamplerSurf; } else { btIndex = nBTIRegularSurf; } stateHeap = renderHal->pStateHeap; // Get Offset to Current Binding Table uint32_t offsetCurrentBTStart = ( stateHeap->iCurSshBufferIndex * stateHeap->dwSshIntanceSize ) + // Points to the Base of Current SSH Buffer Instance ( stateHeap->iBindingTableOffset ) + // Moves the pointer to Base of Array of Binding Tables ( bindingTable * stateHeap->iBindingTableSize ); // Moves the pointer to a Particular Binding Table uint32_t *currentBTStart = ( uint32_t *)( stateHeap->pSshBuffer + offsetCurrentBTStart ); int nEntryIndex = 0; if ( pixelPitch ) { nEntryIndex = (int)((uint32_t*)( state->bti2DIndexTable[ index ].BTITableEntry.samplerBtiEntryPosition ) - currentBTStart); } else { nEntryIndex = (int)((uint32_t*)( state->bti2DIndexTable[ index ].BTITableEntry.regularBtiEntryPosition ) - currentBTStart); } if ( ( nEntryIndex < 0 ) || ( nEntryIndex >= renderHal->StateHeapSettings.iSurfacesPerBT ) ) { nSurfaceEntries = state->bti2DIndexTable[ index ].nPlaneNumber; btIndex = HalCm_GetFreeBindingIndex( state, indexParam, nSurfaceEntries ); // Get Offset to Current Binding Table uint32_t offsetDst = offsetCurrentBTStart + ( btIndex * sizeof( uint32_t ) ); // Move the pointer to correct entry uint32_t *bindingTableEntry = ( uint32_t *)( stateHeap->pSshBuffer + offsetDst ); if ( pixelPitch ) { MOS_SecureMemcpy( bindingTableEntry, sizeof( uint32_t ) * nSurfaceEntries, state->bti2DIndexTable[ index ].BTITableEntry.samplerBtiEntryPosition, sizeof( uint32_t ) * nSurfaceEntries ); } else { MOS_SecureMemcpy( bindingTableEntry, sizeof( uint32_t ) * nSurfaceEntries, state->bti2DIndexTable[ index ].BTITableEntry.regularBtiEntryPosition, sizeof( uint32_t ) * nSurfaceEntries ); } // update index to table if ( pixelPitch ) { state->bti2DIndexTable[ index ].BTI.samplerSurfIndex = btIndex; state->bti2DIndexTable[ index ].BTITableEntry.samplerBtiEntryPosition = bindingTableEntry; } else { state->bti2DIndexTable[ index ].BTI.regularSurfIndex = btIndex; state->bti2DIndexTable[ index ].BTITableEntry.regularBtiEntryPosition = bindingTableEntry; } } // Update the Batch Buffer if ( buffer ) { dst = buffer + argParam->payloadOffset; *( ( uint32_t *)dst ) = btIndex; } eStatus = MOS_STATUS_SUCCESS; goto finish; } CM_CHK_MOSSTATUS_GOTOFINISH( HalCm_GetSurfaceAndRegister( state, &renderHalSurface, CM_ARGUMENT_SURFACE2D, index, pixelPitch ) ); // Setup 2D surface MOS_ZeroMemory(&surfaceParam, sizeof(surfaceParam)); surfaceParam.Type = renderHal->SurfaceTypeDefault; surfaceParam.Boundary = RENDERHAL_SS_BOUNDARY_ORIGINAL; surfaceParam.bVertStride = 0; surfaceParam.bVertStrideOffs = 0; if (!pixelPitch) { surfaceParam.bWidthInDword_UV = true; surfaceParam.bWidthInDword_Y = true; } surfaceParam.isOutput = isRenderTarget(state, index); surfStateParam = &(state->umdSurf2DTable[index].surfaceStateParam[argParam->aliasIndex / state->surfaceArraySize]); if (surfStateParam->width) { surface->dwWidth = surfStateParam->width; } if (surfStateParam->height) { surface->dwHeight = surfStateParam->height; } if (surfStateParam->depth) { surface->dwDepth = surfStateParam->depth; } if (surfStateParam->pitch) { surface->dwPitch= surfStateParam->pitch; } if (surfStateParam->format) { UpdateSurfaceAliasPlaneOffset(surfStateParam, surface); } if (surfStateParam->surfaceXOffset) { surface->YPlaneOffset.iXOffset = surfStateParam->surfaceXOffset; if (surface->Format == Format_NV12) { surface->UPlaneOffset.iXOffset += surfStateParam->surfaceXOffset; } } if (surfStateParam->surfaceYOffset) { surface->YPlaneOffset.iYOffset = surfStateParam->surfaceYOffset; if (surface->Format == Format_NV12) { surface->UPlaneOffset.iYOffset += surfStateParam->surfaceYOffset/2; } } if (surfStateParam->memoryObjectControl) { memObjCtl = surfStateParam->memoryObjectControl; } if(pixelPitch) renderHalSurface.Rotation = state->umdSurf2DTable[index].rotationFlag; //Cache configurations state->cmHalInterface->HwSetSurfaceMemoryObjectControl(memObjCtl, &surfaceParam); // interlace setting HalCm_HwSetSurfaceProperty(state, state->umdSurf2DTable[index].frameType, &surfaceParam); CM_CHK_MOSSTATUS_GOTOFINISH(renderHal->pfnSetupSurfaceState( renderHal, &renderHalSurface, &surfaceParam, &nSurfaceEntries, surfaceEntries, nullptr)); nSurfaceEntries = MOS_MIN( nSurfaceEntries, MHW_MAX_SURFACE_PLANES ); btIndex = HalCm_GetFreeBindingIndex(state, indexParam, nSurfaceEntries); for (i = 0; i < (uint32_t)nSurfaceEntries; i++) { // Bind the surface State CM_CHK_MOSSTATUS_GOTOFINISH(renderHal->pfnBindSurfaceState( renderHal, bindingTable, btIndex + i, surfaceEntries[i])); if ((taskParam->surfEntryInfoArrays.kernelNum !=0) && (taskParam->surfEntryInfoArrays.surfEntryInfosArray != nullptr)) { //GT-Pin CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_GetSurfaceDetails( state, indexParam, btIndex + i, *surface, 0, surfaceEntries[i], tempPlaneIndex, surfaceParam, CM_ARGUMENT_SURFACE2D)); } surfaceEntries[i]->pSurface->dwWidth = state->umdSurf2DTable[index].width; surfaceEntries[i]->pSurface->dwHeight = state->umdSurf2DTable[index].height; } // only update the reuse table for non-aliased surface if ( argParam->aliasCreated == false ) { state->bti2DIndexTable[ index ].nPlaneNumber = nSurfaceEntries; // Get Offset to Current Binding Table stateHeap = renderHal->pStateHeap; offsetSrc = ( stateHeap->iCurSshBufferIndex * stateHeap->dwSshIntanceSize ) + // Points to the Base of Current SSH Buffer Instance ( stateHeap->iBindingTableOffset ) + // Moves the pointer to Base of Array of Binding Tables ( bindingTable * stateHeap->iBindingTableSize ) + // Moves the pointer to a Particular Binding Table ( btIndex * sizeof( uint32_t ) ); // Move the pointer to correct entry if ( pixelPitch ) { state->bti2DIndexTable[ index ].BTI.samplerSurfIndex = btIndex; state->bti2DIndexTable[ index ].BTITableEntry.samplerBtiEntryPosition = stateHeap->pSshBuffer + offsetSrc; } else { state->bti2DIndexTable[ index ].BTI.regularSurfIndex = btIndex; state->bti2DIndexTable[ index ].BTITableEntry.regularBtiEntryPosition = stateHeap->pSshBuffer + offsetSrc; } } // Update the Batch Buffer if (buffer) { dst = buffer + argParam->payloadOffset; *((uint32_t*)dst) = btIndex; } // reset surface height and width surface->dwWidth = state->umdSurf2DTable[index].width; surface->dwHeight = state->umdSurf2DTable[index].height; eStatus = MOS_STATUS_SUCCESS; finish: return eStatus; } MOS_STATUS HalCm_Setup2DSurfaceState( PCM_HAL_STATE state, PCM_HAL_KERNEL_ARG_PARAM argParam, PCM_HAL_INDEX_PARAM indexParam, int32_t bindingTable, uint32_t threadIndex, uint8_t *buffer) { MOS_STATUS eStatus; if (state->cmHalInterface->GetDecompressFlag()) { state->pfnDecompressSurface(state, argParam, threadIndex); } //Binding surface based at the unit of dword CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_Setup2DSurfaceStateBasic( state, argParam, indexParam, bindingTable, threadIndex, false, buffer, false)); eStatus = MOS_STATUS_SUCCESS; finish: return eStatus; } MOS_STATUS HalCm_Setup2DSurfaceSamplerState( PCM_HAL_STATE state, PCM_HAL_KERNEL_ARG_PARAM argParam, PCM_HAL_INDEX_PARAM indexParam, int32_t bindingTable, uint32_t threadIndex, uint8_t *buffer) { MOS_STATUS eStatus; //Binding surface based at the unit of dword CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_Setup2DSurfaceStateBasic( state, argParam, indexParam, bindingTable, threadIndex, true, buffer, false)); eStatus = MOS_STATUS_SUCCESS; finish: return eStatus; } //*----------------------------------------------------------------------------- //| Purpose: Setup 2D surface State //| Returns: Result of the operation //*----------------------------------------------------------------------------- MOS_STATUS HalCm_Setup2DSurfaceUPStateBasic( PCM_HAL_STATE state, PCM_HAL_KERNEL_ARG_PARAM argParam, PCM_HAL_INDEX_PARAM indexParam, int32_t bindingTable, uint32_t threadIndex, uint8_t *buffer, bool pixelPitch) { MOS_STATUS eStatus; RENDERHAL_SURFACE surface; RENDERHAL_SURFACE_STATE_PARAMS surfaceParam; PRENDERHAL_INTERFACE renderHal; PRENDERHAL_SURFACE_STATE_ENTRY surfaceEntries[MHW_MAX_SURFACE_PLANES]; uint8_t *src; uint8_t *dst; int32_t nSurfaceEntries; uint32_t index; uint32_t btIndex; uint16_t memObjCtl; uint32_t i; uint32_t offsetSrc; PRENDERHAL_STATE_HEAP stateHeap; eStatus = MOS_STATUS_UNKNOWN; renderHal = state->renderHal; //GT-PIN PCM_HAL_TASK_PARAM taskParam = state->taskParam; // Get the Index to sampler array from the kernel data CM_ASSERT(argParam->unitSize == sizeof(index)); src = argParam->firstValue + (threadIndex * argParam->unitSize); index = *((uint32_t*)src) & CM_SURFACE_MASK; if (index == CM_NULL_SURFACE) { if (buffer) { dst = buffer + argParam->payloadOffset; *((uint32_t*)dst) = CM_NULL_SURFACE_BINDING_INDEX; } eStatus = MOS_STATUS_SUCCESS; goto finish; } memObjCtl = state->surf2DUPTable[index].memObjCtl; if (!memObjCtl) { memObjCtl = CM_DEFAULT_CACHE_TYPE; } // check to see if the data present for the sampler in the array if (index >= state->cmDeviceParam.max2DSurfaceUPTableSize || (state->surf2DUPTable[index].width == 0)) { eStatus = MOS_STATUS_INVALID_PARAMETER; CM_ASSERTMESSAGE( "Invalid 2D SurfaceUP array index '%d'", index); goto finish; } // Check to see if surface is already assigned if ( pixelPitch ) { btIndex = state->bti2DUPIndexTable[ index ].BTI.samplerSurfIndex; } else { btIndex = state->bti2DUPIndexTable[ index ].BTI.regularSurfIndex; } if ( btIndex == ( unsigned char )CM_INVALID_INDEX ) { uint32_t tempPlaneIndex = 0; // Get Details of 2DUP surface and fill the surface CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_GetSurfaceAndRegister(state, &surface, CM_ARGUMENT_SURFACE2D_UP, index, pixelPitch)); // Setup 2D surface MOS_ZeroMemory(&surfaceParam, sizeof(surfaceParam)); surfaceParam.Type = renderHal->SurfaceTypeDefault; surfaceParam.Boundary = RENDERHAL_SS_BOUNDARY_ORIGINAL; if (!pixelPitch) { surfaceParam.bWidthInDword_UV = true; surfaceParam.bWidthInDword_Y = true; } surfaceParam.isOutput = true; //Cache configurations state->cmHalInterface->HwSetSurfaceMemoryObjectControl(memObjCtl, &surfaceParam); // interlace setting HalCm_HwSetSurfaceProperty(state, state->umdSurf2DTable[index].frameType, &surfaceParam); CM_CHK_MOSSTATUS_GOTOFINISH(renderHal->pfnSetupSurfaceState( renderHal, &surface, &surfaceParam, &nSurfaceEntries, surfaceEntries, nullptr)); //GT-PIN btIndex = HalCm_GetFreeBindingIndex(state, indexParam, nSurfaceEntries); for (i = 0; i < (uint32_t)nSurfaceEntries; i++) { // Bind the surface State CM_CHK_MOSSTATUS_GOTOFINISH(renderHal->pfnBindSurfaceState( renderHal, bindingTable, btIndex + i, surfaceEntries[i])); //GT-Pin if ((taskParam->surfEntryInfoArrays.kernelNum != 0) && (taskParam->surfEntryInfoArrays.surfEntryInfosArray != nullptr)) { CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_GetSurfaceDetails( state, indexParam, btIndex + i, surface.OsSurface, 0, surfaceEntries[i], tempPlaneIndex, surfaceParam, CM_ARGUMENT_SURFACE2D_UP)); } } state->bti2DUPIndexTable[ index ].nPlaneNumber = nSurfaceEntries; stateHeap = renderHal->pStateHeap; offsetSrc = ( stateHeap->iCurSshBufferIndex * stateHeap->dwSshIntanceSize ) + // Points to the Base of Current SSH Buffer Instance ( stateHeap->iBindingTableOffset ) + // Moves the pointer to Base of Array of Binding Tables ( bindingTable * stateHeap->iBindingTableSize ) + // Moves the pointer to a Particular Binding Table ( btIndex * sizeof( uint32_t ) ); // Move the pointer to correct entry if ( pixelPitch ) { state->bti2DUPIndexTable[ index ].BTI.samplerSurfIndex = btIndex; state->bti2DUPIndexTable[ index ].BTITableEntry.samplerBtiEntryPosition = stateHeap->pSshBuffer + offsetSrc; } else { state->bti2DUPIndexTable[ index ].BTI.regularSurfIndex = btIndex; state->bti2DUPIndexTable[ index ].BTITableEntry.regularBtiEntryPosition = stateHeap->pSshBuffer + offsetSrc; } } else { stateHeap = renderHal->pStateHeap; // Get Offset to Current Binding Table uint32_t offsetCurrentBTStart = ( stateHeap->iCurSshBufferIndex * stateHeap->dwSshIntanceSize ) + // Points to the Base of Current SSH Buffer Instance ( stateHeap->iBindingTableOffset ) + // Moves the pointer to Base of Array of Binding Tables ( bindingTable * stateHeap->iBindingTableSize ); // Moves the pointer to a Particular Binding Table uint32_t *currentBTStart = ( uint32_t *)( stateHeap->pSshBuffer + offsetCurrentBTStart ); int nEntryIndex = 0; if ( pixelPitch ) { nEntryIndex = (int) ((uint32_t*)( state->bti2DUPIndexTable[ index ].BTITableEntry.samplerBtiEntryPosition ) - currentBTStart); } else { nEntryIndex = (int) ((uint32_t*)( state->bti2DUPIndexTable[ index ].BTITableEntry.regularBtiEntryPosition ) - currentBTStart); } if ( ( nEntryIndex < 0 ) || ( nEntryIndex >= renderHal->StateHeapSettings.iSurfacesPerBT ) ) { uint32_t tmpSurfaceEntries = state->bti2DUPIndexTable[ index ].nPlaneNumber; btIndex = HalCm_GetFreeBindingIndex( state, indexParam, tmpSurfaceEntries ); // Get Offset to Current Binding Table uint32_t offsetDst = offsetCurrentBTStart + ( btIndex * sizeof( uint32_t ) ); // Move the pointer to correct entry uint32_t *bindingTableEntry = ( uint32_t *)( stateHeap->pSshBuffer + offsetDst ); if ( pixelPitch ) { MOS_SecureMemcpy( bindingTableEntry, sizeof( uint32_t ) * tmpSurfaceEntries, state->bti2DUPIndexTable[ index ].BTITableEntry.samplerBtiEntryPosition, sizeof( uint32_t ) * tmpSurfaceEntries ); } else { MOS_SecureMemcpy( bindingTableEntry, sizeof( uint32_t ) * tmpSurfaceEntries, state->bti2DUPIndexTable[ index ].BTITableEntry.regularBtiEntryPosition, sizeof( uint32_t ) * tmpSurfaceEntries ); } // update index to table if ( pixelPitch ) { state->bti2DUPIndexTable[ index ].BTI.samplerSurfIndex = btIndex; state->bti2DUPIndexTable[ index ].BTITableEntry.samplerBtiEntryPosition = bindingTableEntry; } else { state->bti2DUPIndexTable[ index ].BTI.regularSurfIndex = btIndex; state->bti2DUPIndexTable[ index ].BTITableEntry.regularBtiEntryPosition = bindingTableEntry; } } } // Update the Batch Buffer if (buffer) { dst = buffer + argParam->payloadOffset; *((uint32_t*)dst) = btIndex; } eStatus = MOS_STATUS_SUCCESS; finish: return eStatus; } MOS_STATUS HalCm_Setup2DSurfaceUPState( PCM_HAL_STATE state, PCM_HAL_KERNEL_ARG_PARAM argParam, PCM_HAL_INDEX_PARAM indexParam, int32_t bindingTable, uint32_t threadIndex, uint8_t *buffer) { MOS_STATUS eStatus; //Binding surface based at the unit of dword CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_Setup2DSurfaceUPStateBasic( state, argParam, indexParam, bindingTable, threadIndex, buffer, false)); eStatus = MOS_STATUS_SUCCESS; finish: return eStatus; } MOS_STATUS HalCm_Setup2DSurfaceUPSamplerState( PCM_HAL_STATE state, PCM_HAL_KERNEL_ARG_PARAM argParam, PCM_HAL_INDEX_PARAM indexParam, int32_t bindingTable, uint32_t threadIndex, uint8_t *buffer) { MOS_STATUS eStatus; //Binding surface based at the unit of pixel CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_Setup2DSurfaceUPStateBasic( state, argParam, indexParam, bindingTable, threadIndex, buffer, true)); eStatus = MOS_STATUS_SUCCESS; finish: return eStatus; } MOS_STATUS HalCm_SetupSpecificVmeSurfaceState( PCM_HAL_STATE state, PCM_HAL_INDEX_PARAM indexParam, int32_t bindingTable, uint32_t surfIndex, uint32_t btIndex, uint16_t memObjCtl, uint32_t surfaceStateWidth, uint32_t surfaceStateHeight) { MOS_STATUS eStatus; RENDERHAL_SURFACE surface; int32_t nSurfaceEntries = 0; RENDERHAL_SURFACE_STATE_PARAMS surfaceParam; PRENDERHAL_INTERFACE renderHal; PRENDERHAL_SURFACE_STATE_ENTRY surfaceEntries[MHW_MAX_SURFACE_PLANES]; uint32_t tempPlaneIndex = 0; PMOS_SURFACE mosSurface = nullptr; eStatus = MOS_STATUS_UNKNOWN; renderHal = state->renderHal; nSurfaceEntries = 0; PCM_HAL_TASK_PARAM taskParam = state->taskParam; // Get Details of VME surface and fill the surface CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_GetSurfaceAndRegister(state, &surface, CM_ARGUMENT_VME_STATE, surfIndex, 0)); // Setup 2D surface MOS_ZeroMemory(&surfaceParam, sizeof(surfaceParam)); surfaceParam.Type = renderHal->SurfaceTypeAdvanced; surfaceParam.isOutput = true; surfaceParam.bWidthInDword_Y = false; surfaceParam.bWidthInDword_UV = false; surfaceParam.Boundary = RENDERHAL_SS_BOUNDARY_ORIGINAL; surfaceParam.bVmeUse = true; // Overwrite the width and height if specified if (surfaceStateWidth && surfaceStateHeight) { mosSurface = &surface.OsSurface; if (surfaceStateWidth > mosSurface->dwWidth || surfaceStateHeight > mosSurface->dwHeight) { CM_ASSERTMESSAGE("Error: VME surface state's resolution is larger than the original surface."); eStatus = MOS_STATUS_INVALID_PARAMETER; goto finish; } mosSurface->dwWidth = surfaceStateWidth; mosSurface->dwHeight = surfaceStateHeight; } //Cache configurations state->cmHalInterface->HwSetSurfaceMemoryObjectControl(memObjCtl, &surfaceParam); CM_CHK_MOSSTATUS_GOTOFINISH(renderHal->pfnSetupSurfaceState( renderHal, &surface, &surfaceParam, &nSurfaceEntries, surfaceEntries, nullptr)); CM_ASSERT(nSurfaceEntries == 1); { // Bind the surface State CM_CHK_MOSSTATUS_GOTOFINISH(renderHal->pfnBindSurfaceState( renderHal, bindingTable, btIndex, surfaceEntries[0])); if ((taskParam->surfEntryInfoArrays.kernelNum != 0) && (taskParam->surfEntryInfoArrays.surfEntryInfosArray != nullptr)) { CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_GetSurfaceDetails( state, indexParam, btIndex, surface.OsSurface, 0, surfaceEntries[0], tempPlaneIndex, surfaceParam, CM_ARGUMENT_SURFACE2D)); } } state->bti2DIndexTable[ surfIndex ].BTI.vmeSurfIndex = btIndex; eStatus = MOS_STATUS_SUCCESS; finish: return eStatus; } //*----------------------------------------------------------------------------- //| Purpose: Setup VME surface State //| Returns: Result of the operation //*----------------------------------------------------------------------------- MOS_STATUS HalCm_SetupVmeSurfaceState( PCM_HAL_STATE state, PCM_HAL_KERNEL_ARG_PARAM argParam, PCM_HAL_INDEX_PARAM indexParam, int32_t bindingTable, uint32_t threadIndex, uint8_t *buffer) { MOS_STATUS eStatus; PRENDERHAL_INTERFACE renderHal; PCM_HAL_VME_ARG_VALUE vmeSrc; uint8_t *dst; uint32_t index[CM_MAX_VME_BINDING_INDEX_1]; uint16_t memObjCtl[CM_MAX_VME_BINDING_INDEX_1]; uint32_t fwSurfCount = 0; uint32_t bwSurfCount = 0; bool alreadyBind = true; uint32_t surfPairNum; uint32_t idx; uint32_t curBTIndex; uint32_t btIndex; uint32_t surfaceStateWidth = 0; uint32_t surfaceStateHeight = 0; uint32_t *fPtr = nullptr; uint32_t *bPtr = nullptr; uint32_t *refSurfaces = nullptr; eStatus = MOS_STATUS_UNKNOWN; renderHal = state->renderHal; btIndex = 0; MOS_ZeroMemory(memObjCtl, CM_MAX_VME_BINDING_INDEX_1*sizeof(uint16_t)); MOS_ZeroMemory(index, CM_MAX_VME_BINDING_INDEX_1*sizeof(uint32_t)); CM_ASSERT(argParam->unitSize <= sizeof(uint32_t)*(CM_MAX_VME_BINDING_INDEX_1 + 2)); CM_ASSERT(threadIndex == 0); // VME surface is not allowed in thread arg vmeSrc = (PCM_HAL_VME_ARG_VALUE)argParam->firstValue; fwSurfCount = vmeSrc->fwRefNum; bwSurfCount = vmeSrc->bwRefNum; refSurfaces = findRefInVmeArg(vmeSrc); index[0] = vmeSrc->curSurface & CM_SURFACE_MASK; // check to see if index[0] is valid if (index[0] == CM_NULL_SURFACE) { if (buffer) { dst = buffer + argParam->payloadOffset; *((uint32_t*)dst) = CM_NULL_SURFACE_BINDING_INDEX; } eStatus = MOS_STATUS_SUCCESS; goto finish; } if (index[0] >= state->cmDeviceParam.max2DSurfaceTableSize || Mos_ResourceIsNull(&state->umdSurf2DTable[index[0]].osResource)) { eStatus = MOS_STATUS_INVALID_PARAMETER; CM_ASSERTMESSAGE( "Invalid 2D surface array index '%d'", index[0]); goto finish; } memObjCtl[0] = state->umdSurf2DTable[index[0]].memObjCtl; if (!memObjCtl[0]) { memObjCtl[0] = CM_DEFAULT_CACHE_TYPE; } for (idx = 0; idx < (vmeSrc->fwRefNum + vmeSrc->bwRefNum); idx++) { index[idx + 1] = refSurfaces[idx] & CM_SURFACE_MASK; memObjCtl[idx + 1] = state->umdSurf2DTable[index[idx + 1]].memObjCtl; if (!memObjCtl[idx + 1]) { memObjCtl[idx + 1] = CM_DEFAULT_CACHE_TYPE; } } surfaceStateWidth = vmeSrc->surfStateParam.surfaceStateWidth; surfaceStateHeight = vmeSrc->surfStateParam.surfaceStateHeight; fPtr = index + 1; bPtr = index + 1 + fwSurfCount; //Max surface pair number surfPairNum = fwSurfCount > bwSurfCount ? fwSurfCount : bwSurfCount; btIndex = curBTIndex = HalCm_GetFreeBindingIndex(state, indexParam, surfPairNum*2 + 1); HalCm_SetupSpecificVmeSurfaceState(state, indexParam, bindingTable, index[0], curBTIndex, memObjCtl[0], surfaceStateWidth, surfaceStateHeight); curBTIndex++; //Setup surface states interleavely for backward and forward surfaces pairs. for (idx = 0; idx < surfPairNum; idx++) { if (idx < fwSurfCount) { HalCm_SetupSpecificVmeSurfaceState(state, indexParam, bindingTable, fPtr[idx], curBTIndex, memObjCtl[idx + 1], surfaceStateWidth, surfaceStateHeight); } curBTIndex++; if (idx < bwSurfCount) { HalCm_SetupSpecificVmeSurfaceState(state, indexParam, bindingTable, bPtr[idx], curBTIndex, memObjCtl[idx+ 1 + fwSurfCount], surfaceStateWidth, surfaceStateHeight); } curBTIndex++; } // Update the Batch Buffer if (buffer) { dst = buffer + argParam->payloadOffset; *((uint32_t*)dst) = btIndex; } eStatus = MOS_STATUS_SUCCESS; finish: return eStatus; } static bool UpdateMosSurfaceFromAliasState(CM_HAL_STATE *state, CM_HAL_KERNEL_ARG_PARAM *argParam, uint32_t surface_index, MOS_SURFACE *surface) { uint32_t surface_state_index = argParam->aliasIndex/state->surfaceArraySize; const CM_HAL_SURFACE2D_SURFACE_STATE_PARAM &surface_state_param = state->umdSurf2DTable[surface_index].surfaceStateParam[ surface_state_index]; if (surface_state_param.width) { surface->dwWidth = surface_state_param.width; } if (surface_state_param.height) { surface->dwHeight = surface_state_param.height; } if (surface_state_param.depth) { surface->dwDepth = surface_state_param.depth; } if (surface_state_param.pitch) { surface->dwPitch= surface_state_param.pitch; } if (surface_state_param.format) { surface->Format = static_cast(surface_state_param.format); } if (surface_state_param.surfaceXOffset) { surface->YPlaneOffset.iXOffset = surface_state_param.surfaceXOffset; } if (surface_state_param.surfaceYOffset) { surface->YPlaneOffset.iYOffset = surface_state_param.surfaceYOffset; } if (surface_state_param.surfaceOffset) { surface->dwOffset = surface_state_param.surfaceOffset; } return true; } //*----------------------------------------------------------------------------- //| Purpose: Setup VME surface State //| Returns: Result of the operation //*----------------------------------------------------------------------------- MOS_STATUS HalCm_SetupSampler8x8SurfaceState( PCM_HAL_STATE state, PCM_HAL_KERNEL_ARG_PARAM argParam, PCM_HAL_INDEX_PARAM indexParam, int32_t bindingTable, uint32_t threadIndex, uint8_t *buffer) { MOS_STATUS eStatus; RENDERHAL_SURFACE surface; RENDERHAL_SURFACE_STATE_PARAMS surfaceParam; PRENDERHAL_INTERFACE renderHal; PRENDERHAL_SURFACE_STATE_ENTRY surfaceEntries[MHW_MAX_SURFACE_PLANES]; uint8_t *src; uint8_t *dst; int32_t nSurfaceEntries; uint32_t index; uint16_t memObjCtl; int32_t i; uint32_t btIndex; uint32_t tempPlaneIndex = 0; uint32_t offsetSrc; PRENDERHAL_STATE_HEAP stateHeap; eStatus = MOS_STATUS_UNKNOWN; renderHal = state->renderHal; PCM_HAL_TASK_PARAM taskParam = state->taskParam; nSurfaceEntries = 0; CM_ASSERT(argParam->unitSize == sizeof(uint32_t)); src = argParam->firstValue + (threadIndex * argParam->unitSize); index = *((uint32_t*)src) & CM_SURFACE_MASK; if (index == CM_NULL_SURFACE) { if (buffer) { dst = buffer + argParam->payloadOffset; *((uint32_t*)dst) = CM_NULL_SURFACE_BINDING_INDEX; } eStatus = MOS_STATUS_SUCCESS; goto finish; } memObjCtl = state->umdSurf2DTable[index].memObjCtl; if (!memObjCtl) { memObjCtl = CM_DEFAULT_CACHE_TYPE; } // check to see if index is valid if (index >= state->cmDeviceParam.max2DSurfaceTableSize || Mos_ResourceIsNull(&state->umdSurf2DTable[index].osResource)) { eStatus = MOS_STATUS_INVALID_PARAMETER; CM_ASSERTMESSAGE( "Invalid 2D surface array index '%d'", index); goto finish; } renderHal->bEnableP010SinglePass = state->cmHalInterface->IsP010SinglePassSupported(); btIndex = state->bti2DIndexTable[ index ].BTI.sampler8x8SurfIndex; if (btIndex == ( unsigned char )CM_INVALID_INDEX || argParam->aliasCreated) { // Get Details of Sampler8x8 surface and fill the surface CM_CHK_MOSSTATUS_GOTOFINISH( HalCm_GetSurfaceAndRegister( state, &surface, argParam->kind, index, 0 ) ); // Setup surface MOS_ZeroMemory( &surfaceParam, sizeof( surfaceParam ) ); surfaceParam.Type = renderHal->SurfaceTypeAdvanced; surfaceParam.isOutput = true; surfaceParam.bWidthInDword_Y = false; surfaceParam.bWidthInDword_UV = false; surfaceParam.Boundary = RENDERHAL_SS_BOUNDARY_ORIGINAL; surfaceParam.bVASurface = ( argParam->kind == CM_ARGUMENT_SURFACE_SAMPLER8X8_VA ) ? 1 : 0; surfaceParam.AddressControl = argParam->nCustomValue; UpdateMosSurfaceFromAliasState(state, argParam, index, &surface.OsSurface); //Set memory object control state->cmHalInterface->HwSetSurfaceMemoryObjectControl(memObjCtl, &surfaceParam); surface.Rotation = state->umdSurf2DTable[index].rotationFlag; surface.ChromaSiting = state->umdSurf2DTable[index].chromaSiting; surface.ScalingMode = RENDERHAL_SCALING_AVS; nSurfaceEntries = 0; // interlace setting HalCm_HwSetSurfaceProperty(state, state->umdSurf2DTable[index].frameType, &surfaceParam); CM_CHK_MOSSTATUS_GOTOFINISH( renderHal->pfnSetupSurfaceState( renderHal, &surface, &surfaceParam, &nSurfaceEntries, surfaceEntries, nullptr ) ); CM_ASSERT( nSurfaceEntries == 1 ); btIndex = HalCm_GetFreeBindingIndex( state, indexParam, nSurfaceEntries ); for ( i = 0; i < nSurfaceEntries; i++ ) { // Bind the surface State CM_CHK_MOSSTATUS_GOTOFINISH( renderHal->pfnBindSurfaceState( renderHal, bindingTable, btIndex + i, surfaceEntries[ i ] ) ); if ( ( taskParam->surfEntryInfoArrays.kernelNum != 0 ) && ( taskParam->surfEntryInfoArrays.surfEntryInfosArray != nullptr ) ) { CM_CHK_MOSSTATUS_GOTOFINISH( HalCm_GetSurfaceDetails( state, indexParam, btIndex + i, surface.OsSurface, 0, surfaceEntries[ i ], tempPlaneIndex, surfaceParam, CM_ARGUMENT_SURFACE2D ) ); } } stateHeap = renderHal->pStateHeap; offsetSrc = ( stateHeap->iCurSshBufferIndex * stateHeap->dwSshIntanceSize ) + // Points to the Base of Current SSH Buffer Instance ( stateHeap->iBindingTableOffset ) + // Moves the pointer to Base of Array of Binding Tables ( bindingTable * stateHeap->iBindingTableSize ) + // Moves the pointer to a Particular Binding Table ( btIndex * sizeof( uint32_t ) ); // Move the pointer to correct entry state->bti2DIndexTable[ index ].nPlaneNumber = nSurfaceEntries; state->bti2DIndexTable[ index ].BTITableEntry.sampler8x8BtiEntryPosition = stateHeap->pSshBuffer + offsetSrc; state->bti2DIndexTable[ index ].BTI.sampler8x8SurfIndex = btIndex; } else { stateHeap = renderHal->pStateHeap; // Get Offset to Current Binding Table uint32_t offsetCurrentBTStart = ( stateHeap->iCurSshBufferIndex * stateHeap->dwSshIntanceSize ) + // Points to the Base of Current SSH Buffer Instance ( stateHeap->iBindingTableOffset ) + // Moves the pointer to Base of Array of Binding Tables ( bindingTable * stateHeap->iBindingTableSize ); // Moves the pointer to a Particular Binding Table uint32_t *currentBTStart = ( uint32_t *)( stateHeap->pSshBuffer + offsetCurrentBTStart ); int nEntryIndex = 0; nEntryIndex = ( int )( ( uint32_t *)( state->bti2DIndexTable[ index ].BTITableEntry.sampler8x8BtiEntryPosition ) - currentBTStart ); if ( ( nEntryIndex < 0 ) || ( nEntryIndex >= renderHal->StateHeapSettings.iSurfacesPerBT ) ) { uint32_t tmpSurfaceEntries = state->bti2DIndexTable[ index ].nPlaneNumber; btIndex = HalCm_GetFreeBindingIndex( state, indexParam, tmpSurfaceEntries ); // Get Offset to Current Binding Table uint32_t offsetDst = offsetCurrentBTStart + ( btIndex * sizeof( uint32_t ) ); // Move the pointer to correct entry uint32_t *bindingTableEntry = ( uint32_t *)( stateHeap->pSshBuffer + offsetDst ); MOS_SecureMemcpy( bindingTableEntry, sizeof( uint32_t ) * tmpSurfaceEntries, state->bti2DIndexTable[ index ].BTITableEntry.sampler8x8BtiEntryPosition, sizeof( uint32_t ) * tmpSurfaceEntries ); // update index to table state->bti2DIndexTable[ index ].BTI.sampler8x8SurfIndex = btIndex; state->bti2DIndexTable[ index ].BTITableEntry.sampler8x8BtiEntryPosition = bindingTableEntry; } } // Update the Batch Buffer if ( buffer ) { dst = buffer + argParam->payloadOffset; *( ( uint32_t *)dst ) = state->bti2DIndexTable[ index ].BTI.sampler8x8SurfIndex; } eStatus = MOS_STATUS_SUCCESS; finish: renderHal->bEnableP010SinglePass = false; return eStatus; } //*----------------------------------------------------------------------------- //| Purpose: Setup State Buffer surface State //| Returns: Result of the operation //*----------------------------------------------------------------------------- MOS_STATUS HalCm_SetupStateBufferSurfaceState( PCM_HAL_STATE state, PCM_HAL_KERNEL_ARG_PARAM argParam, PCM_HAL_INDEX_PARAM indexParam, int32_t bindingTable, uint32_t threadIndex, uint8_t *buffer ) { MOS_STATUS eStatus = MOS_STATUS_SUCCESS; PRENDERHAL_INTERFACE renderHal; RENDERHAL_SURFACE_STATE_PARAMS surfaceParam; RENDERHAL_SURFACE renderhalSurface; PRENDERHAL_SURFACE_STATE_ENTRY surfaceEntry; uint32_t btIndex; CM_SURFACE_BTI_INFO surfBTIInfo; uint16_t memObjCtl; state->cmHalInterface->GetHwSurfaceBTIInfo( &surfBTIInfo ); uint32_t surfIndex = reinterpret_cast< uint32_t *>( argParam->firstValue )[ 0 ]; surfIndex = surfIndex & CM_SURFACE_MASK; memObjCtl = state->bufferTable[ surfIndex ].memObjCtl; btIndex = HalCm_GetFreeBindingIndex( state, indexParam, 1 ); renderHal = state->renderHal; MOS_ZeroMemory( &renderhalSurface, sizeof( renderhalSurface ) ); // Get Details of Sampler8x8 surface and fill the surface CM_CHK_MOSSTATUS_GOTOFINISH( HalCm_GetSurfaceAndRegister( state, &renderhalSurface, argParam->kind, surfIndex, 0 ) ); MOS_ZeroMemory( &surfaceParam, sizeof( surfaceParam ) ); // Set the isOutput by default surfaceParam.isOutput = true; //Cache configurations default state->cmHalInterface->HwSetSurfaceMemoryObjectControl( memObjCtl, &surfaceParam ); // Setup Buffer surface CM_CHK_MOSSTATUS_GOTOFINISH( renderHal->pfnSetupBufferSurfaceState( renderHal, &renderhalSurface, &surfaceParam, &surfaceEntry ) ); // Bind the surface State CM_ASSERT( ( ( int32_t )btIndex ) < renderHal->StateHeapSettings.iSurfacesPerBT + surfBTIInfo.normalSurfaceStart ); CM_CHK_MOSSTATUS_GOTOFINISH( renderHal->pfnBindSurfaceState( renderHal, bindingTable, btIndex, surfaceEntry ) ); if ( buffer ) { *( ( uint32_t *)( buffer + argParam->payloadOffset ) ) = btIndex; } finish: return eStatus; } //------------------------------------------------------------------------------ //| Purpose: Get usr defined threadcount / threadgroup //| Returns: Result of the operation //------------------------------------------------------------------------------ MOS_STATUS HalCm_GetMaxThreadCountPerThreadGroup( PCM_HAL_STATE state, // [in] Pointer to CM State uint32_t *threadsPerThreadGroup) // [out] Pointer to threadsPerThreadGroup { MOS_STATUS eStatus = MOS_STATUS_SUCCESS; CM_PLATFORM_INFO platformInfo; MOS_ZeroMemory(&platformInfo, sizeof(CM_PLATFORM_INFO)); CM_CHK_MOSSTATUS_GOTOFINISH( state->pfnGetPlatformInfo( state, &platformInfo, false) ); if (platformInfo.numMaxEUsPerPool) { *threadsPerThreadGroup = (platformInfo.numHWThreadsPerEU) * (platformInfo.numMaxEUsPerPool); } else { *threadsPerThreadGroup = (platformInfo.numHWThreadsPerEU) * (platformInfo.numEUsPerSubSlice); } finish: return eStatus; } //*----------------------------------------------------------------------------- //| Purpose: Decodes hints to get number and size of kernel groups //| Returns: Result of the operation //*----------------------------------------------------------------------------- MOS_STATUS HalCm_GetNumKernelsPerGroup( uint8_t hintsBits, uint32_t numKernels, uint32_t *numKernelsPerGroup, uint32_t *numKernelGroups, uint32_t *remapKernelToGroup, uint32_t *remapGroupToKernel ) { MOS_STATUS eStatus = MOS_STATUS_SUCCESS; uint32_t currGrp = 0; uint32_t i = 0; // first group at least has one kernel numKernelsPerGroup[currGrp]++; remapGroupToKernel[currGrp] = 0; for( i = 0; i < numKernels - 1; ++i ) { if( (hintsBits & CM_HINTS_LEASTBIT_MASK) == CM_HINTS_LEASTBIT_MASK ) { currGrp++; *numKernelGroups = *numKernelGroups + 1; remapGroupToKernel[currGrp] = i + 1; } numKernelsPerGroup[currGrp]++; hintsBits >>= 1; remapKernelToGroup[i+1] = currGrp; } return eStatus; } //*----------------------------------------------------------------------------- //| Purpose: Gets information about max parallelism graphs //| numThreadsOnSides based on formula to sum 1 to n: (n(n+1))/2 //| Returns: Result of the operation //*----------------------------------------------------------------------------- MOS_STATUS HalCm_GetParallelGraphInfo( uint32_t maximum, uint32_t numThreads, uint32_t width, uint32_t height, PCM_HAL_PARALLELISM_GRAPH_INFO graphInfo, CM_DEPENDENCY_PATTERN pattern, bool noDependencyCase) { MOS_STATUS eStatus = MOS_STATUS_SUCCESS; uint32_t numThreadsOnSides = 0; uint32_t numMaxRepeat = 0; uint32_t numSteps = 0; switch( pattern ) { case CM_NONE_DEPENDENCY: if (noDependencyCase) { maximum = 1; numMaxRepeat = width * height; numSteps = width * height; } // do nothing will depend on other kernels break; case CM_VERTICAL_WAVE: numMaxRepeat = width; numSteps = width; break; case CM_HORIZONTAL_WAVE: numMaxRepeat = height; numSteps = height; break; case CM_WAVEFRONT: numThreadsOnSides = ( maximum - 1 ) * maximum; numMaxRepeat = (numThreads - numThreadsOnSides ) / maximum; numSteps = ( maximum - 1) * 2 + numMaxRepeat; break; case CM_WAVEFRONT26: numThreadsOnSides = ( maximum - 1 ) * maximum * 2; numMaxRepeat = (numThreads - numThreadsOnSides ) / maximum; numSteps = ( (maximum - 1) * 2 ) * 2 + numMaxRepeat; break; case CM_WAVEFRONT26Z: // do nothing already set outside of this function break; default: eStatus = MOS_STATUS_INVALID_PARAMETER; CM_ASSERTMESSAGE("Unsupported dependency pattern for EnqueueWithHints"); goto finish; } graphInfo->maxParallelism = maximum; graphInfo->numMaxRepeat = numMaxRepeat; graphInfo->numSteps = numSteps; finish: return eStatus; } //*----------------------------------------------------------------------------- //| Purpose: Sets dispatch pattern based on max parallelism for media objects //| Returns: Result of the operation //*----------------------------------------------------------------------------- MOS_STATUS HalCm_SetDispatchPattern( CM_HAL_PARALLELISM_GRAPH_INFO graphInfo, CM_DEPENDENCY_PATTERN pattern, uint32_t *dispatchFreq ) { MOS_STATUS eStatus = MOS_STATUS_SUCCESS; uint32_t i = 0; uint32_t j = 0; uint32_t k = 0; switch( pattern ) { case CM_NONE_DEPENDENCY: break; case CM_HORIZONTAL_WAVE: case CM_VERTICAL_WAVE: for( i = 0; i < graphInfo.numSteps; ++i ) { dispatchFreq[i] = graphInfo.maxParallelism; } break; case CM_WAVEFRONT: for( i = 1; i < graphInfo.maxParallelism; ++i ) { dispatchFreq[i-1] = i; } for( j = 0; j < graphInfo.numMaxRepeat; ++i, ++j ) { dispatchFreq[i-1] = graphInfo.maxParallelism; } for( j = graphInfo.maxParallelism - 1; i <= graphInfo.numSteps; ++i, --j ) { dispatchFreq[i-1] = j; } break; case CM_WAVEFRONT26: for( i = 1, j = 0; i < graphInfo.maxParallelism; ++i, j +=2 ) { dispatchFreq[j] = i; dispatchFreq[j+1] = i; } for( k = 0; k < graphInfo.numMaxRepeat; ++k, ++j) { dispatchFreq[j] = graphInfo.maxParallelism; } for( i = graphInfo.maxParallelism - 1; j < graphInfo.numSteps; j +=2, --i ) { dispatchFreq[j] = i; dispatchFreq[j+1] = i; } break; case CM_WAVEFRONT26Z: break; default: eStatus = MOS_STATUS_INVALID_PARAMETER; CM_ASSERTMESSAGE("Unsupported dependency pattern for EnqueueWithHints"); goto finish; } finish: return eStatus; } //*----------------------------------------------------------------------------- //| Purpose: Sets dispatch frequency for kernel group based on number of steps //| Returns: Result of the operation //*----------------------------------------------------------------------------- MOS_STATUS HalCm_SetKernelGrpFreqDispatch( PCM_HAL_PARALLELISM_GRAPH_INFO graphInfo, PCM_HAL_KERNEL_GROUP_INFO groupInfo, uint32_t numKernelGroups, uint32_t *minSteps) { MOS_STATUS eStatus = MOS_STATUS_SUCCESS; uint32_t i = 0; uint32_t j = 0; uint32_t tmpSteps = 0; uint32_t kerIndex = 0; for( i = 0; i < numKernelGroups; ++i) { for( j = 0; j < groupInfo[i].numKernelsInGroup; ++j ) { tmpSteps += graphInfo[kerIndex].numSteps; kerIndex++; } if ( tmpSteps ) { *minSteps = MOS_MIN(*minSteps, tmpSteps); groupInfo[i].numStepsInGrp = tmpSteps; } tmpSteps = 0; } for( i = 0; i < numKernelGroups; ++i ) { groupInfo[i].freqDispatch = (uint32_t)ceil( (groupInfo[i].numStepsInGrp / (double)*minSteps) ); } return eStatus; } //*----------------------------------------------------------------------------- //| Purpose: Sets dispatch pattern for kernel with no dependency based on //| the minimum number of steps calculated from kernels with dependency //| Returns: Result of the operation //*----------------------------------------------------------------------------- MOS_STATUS HalCm_SetNoDependKernelDispatchPattern( uint32_t numThreads, uint32_t minSteps, uint32_t *dispatchFreq) { MOS_STATUS eStatus = MOS_STATUS_SUCCESS; uint32_t i = 0; uint32_t numEachStep = 0; uint32_t total = 0; numEachStep = numThreads / minSteps; for( i = 0; i < minSteps; ++i ) { dispatchFreq[i] = numEachStep; total += numEachStep; } while( total != numThreads ) { // dispatch more at beginning i = 0; dispatchFreq[i]++; total++; i++; } return eStatus; } MOS_STATUS HalCm_FinishStatesForKernel( PCM_HAL_STATE state, // [in] Pointer to CM State PRENDERHAL_MEDIA_STATE mediaState, PMHW_BATCH_BUFFER batchBuffer, // [in] Pointer to Batch Buffer int32_t taskId, // [in] Task ID PCM_HAL_KERNEL_PARAM kernelParam, int32_t kernelIndex, PCM_HAL_INDEX_PARAM indexParam, int32_t bindingTable, int32_t mediaID, PRENDERHAL_KRN_ALLOCATION krnAllocation ) { MOS_STATUS eStatus = MOS_STATUS_SUCCESS; PCM_HAL_TASK_PARAM taskParam = state->taskParam; PRENDERHAL_INTERFACE renderHal = state->renderHal; PCM_HAL_WALKER_PARAMS mediaWalkerParams = &kernelParam->walkerParams; PCM_GPGPU_WALKER_PARAMS perKernelGpGpuWalkerParams = &kernelParam->gpgpuWalkerParams; PCM_HAL_SCOREBOARD threadCoordinates = nullptr; PCM_HAL_MASK_AND_RESET dependencyMask = nullptr; bool enableThreadSpace = false; bool enableKernelThreadSpace = false; PCM_HAL_SCOREBOARD kernelThreadCoordinates = nullptr; UNUSED(taskId); MHW_MEDIA_OBJECT_PARAMS mediaObjectParam; PCM_HAL_KERNEL_ARG_PARAM argParam; MHW_PIPE_CONTROL_PARAMS pipeControlParam; uint32_t i; uint32_t hdrSize; uint32_t aIndex; uint32_t tIndex; uint32_t index; //GT-PIN taskParam->curKernelIndex = kernelIndex; CmSafeMemSet(&mediaObjectParam, 0, sizeof(MHW_MEDIA_OBJECT_PARAMS)); if (perKernelGpGpuWalkerParams->gpgpuEnabled) { // GPGPU_WALKER, just update ID here. other fields are already filled. perKernelGpGpuWalkerParams->interfaceDescriptorOffset = mediaID;// mediaObjectParam.dwInterfaceDescriptorOffset; } else if (mediaWalkerParams->cmWalkerEnable) { // Media walker, just update ID here. other fields are already filled. mediaWalkerParams->interfaceDescriptorOffset = mediaID; } else { // MEDIA_OBJECT mediaObjectParam.dwInterfaceDescriptorOffset = mediaID; hdrSize = renderHal->pHwSizes->dwSizeMediaObjectHeaderCmd; if (kernelParam->indirectDataParam.indirectDataSize) { mediaObjectParam.dwInlineDataSize = 0; } else { mediaObjectParam.dwInlineDataSize = MOS_MAX(kernelParam->payloadSize, 4); } if (taskParam->threadCoordinates) { threadCoordinates = taskParam->threadCoordinates[kernelIndex]; if (threadCoordinates) { enableThreadSpace = true; } } else if (kernelParam->kernelThreadSpaceParam.threadCoordinates) { kernelThreadCoordinates = kernelParam->kernelThreadSpaceParam.threadCoordinates; if (kernelThreadCoordinates) { enableKernelThreadSpace = true; } } if (taskParam->dependencyMasks) { dependencyMask = taskParam->dependencyMasks[kernelIndex]; } CM_CHK_NULL_GOTOFINISH_MOSERROR( batchBuffer ); uint8_t inlineData[CM_MAX_THREAD_PAYLOAD_SIZE]; uint8_t *cmdInline = inlineData; uint32_t cmdSize = mediaObjectParam.dwInlineDataSize + hdrSize; // Setup states for arguments and threads if (((PCM_HAL_BB_ARGS)batchBuffer->pPrivateData)->refCount > 1) { uint8_t *bBuffer = batchBuffer->pData + batchBuffer->iCurrent; for (aIndex = 0; aIndex < kernelParam->numArgs; aIndex++) { argParam = &kernelParam->argParams[aIndex]; if ((kernelParam->cmFlags & CM_KERNEL_FLAGS_CURBE) && !argParam->perThread) { continue; } for (tIndex = 0; tIndex < kernelParam->numThreads; tIndex++) { index = tIndex * argParam->perThread; //----------------------------------------------------- CM_ASSERT(argParam->payloadOffset < kernelParam->payloadSize); //----------------------------------------------------- switch(argParam->kind) { case CM_ARGUMENT_GENERAL: break; case CM_ARGUMENT_SAMPLER: CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_SetupSamplerState( state, kernelParam, argParam, indexParam, mediaID, index, nullptr)); break; case CM_ARGUMENT_SURFACEBUFFER: CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_SetupBufferSurfaceState( state, argParam, indexParam, bindingTable, -1, index, nullptr)); break; case CM_ARGUMENT_SURFACE2D_UP: CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_Setup2DSurfaceUPState( state, argParam, indexParam, bindingTable, index, nullptr)); break; case CM_ARGUMENT_SURFACE2DUP_SAMPLER: CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_Setup2DSurfaceUPSamplerState( state, argParam, indexParam, bindingTable, index, nullptr)); break; case CM_ARGUMENT_SURFACE2D_SAMPLER: CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_Setup2DSurfaceSamplerState( state, argParam, indexParam, bindingTable, 0, nullptr)); break; case CM_ARGUMENT_SURFACE2D: CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_Setup2DSurfaceState( state, argParam, indexParam, bindingTable, index, nullptr)); break; case CM_ARGUMENT_SURFACE3D: CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_Setup3DSurfaceState( state, argParam, indexParam, bindingTable, index, nullptr)); break; case CM_ARGUMENT_SURFACE_VME: CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_SetupVmeSurfaceState( state, argParam, indexParam, bindingTable, 0, nullptr)); break; case CM_ARGUMENT_SURFACE_SAMPLER8X8_AVS: case CM_ARGUMENT_SURFACE_SAMPLER8X8_VA: CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_SetupSampler8x8SurfaceState( state, argParam, indexParam, bindingTable, 0, nullptr)); break; default: eStatus = MOS_STATUS_INVALID_PARAMETER; CM_ASSERTMESSAGE( "Argument kind '%d' is not supported", argParam->kind); goto finish; } } if( dependencyMask ) { if( dependencyMask[tIndex].resetMask == CM_RESET_DEPENDENCY_MASK ) { MOS_SecureMemcpy(bBuffer + (CM_SCOREBOARD_MASK_POS_IN_MEDIA_OBJECT_CMD*sizeof(uint32_t)), sizeof(uint8_t), &dependencyMask[tIndex].mask, sizeof(uint8_t)); } } batchBuffer->iCurrent += cmdSize; bBuffer += cmdSize; } } else { //Insert synchronization if needed (PIPE_CONTROL) // 1. synchronization is set // 2. the next kernel has dependency pattern if((kernelIndex > 0) && ((taskParam->syncBitmap & ((uint64_t)1 << (kernelIndex-1))) || (kernelParam->kernelThreadSpaceParam.patternType != CM_NONE_DEPENDENCY))) { pipeControlParam = g_cRenderHal_InitPipeControlParams; pipeControlParam.presDest = nullptr; pipeControlParam.dwFlushMode = MHW_FLUSH_CUSTOM; // Use custom flags pipeControlParam.dwPostSyncOp = MHW_FLUSH_NOWRITE; pipeControlParam.bDisableCSStall = false; pipeControlParam.bTlbInvalidate = false; pipeControlParam.bFlushRenderTargetCache = true; pipeControlParam.bInvalidateTextureCache = true; CM_CHK_MOSSTATUS_RETURN(renderHal->pMhwMiInterface->AddPipeControl(nullptr, batchBuffer, &pipeControlParam)); } uint8_t *bBuffer = batchBuffer->pData + batchBuffer->iCurrent; for (tIndex = 0; tIndex < kernelParam->numThreads; tIndex++) { if (enableThreadSpace) { mediaObjectParam.VfeScoreboard.ScoreboardEnable = (state->scoreboardParams.ScoreboardMask==0) ? 0:1; mediaObjectParam.VfeScoreboard.Value[0] = threadCoordinates[tIndex].x; mediaObjectParam.VfeScoreboard.Value[1] = threadCoordinates[tIndex].y; mediaObjectParam.VfeScoreboard.ScoreboardColor = threadCoordinates[tIndex].color; mediaObjectParam.dwSliceDestinationSelect = threadCoordinates[tIndex].sliceSelect; mediaObjectParam.dwHalfSliceDestinationSelect = threadCoordinates[tIndex].subSliceSelect; if( !dependencyMask ) { mediaObjectParam.VfeScoreboard.ScoreboardMask = (1 << state->scoreboardParams.ScoreboardMask)-1; } else { mediaObjectParam.VfeScoreboard.ScoreboardMask = dependencyMask[tIndex].mask; } } else if (enableKernelThreadSpace) { mediaObjectParam.VfeScoreboard.ScoreboardEnable = (state->scoreboardParams.ScoreboardMask == 0) ? 0 : 1; mediaObjectParam.VfeScoreboard.Value[0] = kernelThreadCoordinates[tIndex].x; mediaObjectParam.VfeScoreboard.Value[1] = kernelThreadCoordinates[tIndex].y; mediaObjectParam.VfeScoreboard.ScoreboardColor = kernelThreadCoordinates[tIndex].color; mediaObjectParam.dwSliceDestinationSelect = kernelThreadCoordinates[tIndex].sliceSelect; mediaObjectParam.dwHalfSliceDestinationSelect = kernelThreadCoordinates[tIndex].subSliceSelect; if (!dependencyMask) { mediaObjectParam.VfeScoreboard.ScoreboardMask = (1 << state->scoreboardParams.ScoreboardMask) - 1; } else { mediaObjectParam.VfeScoreboard.ScoreboardMask = dependencyMask[tIndex].mask; } } else { mediaObjectParam.VfeScoreboard.Value[0] = tIndex % taskParam->threadSpaceWidth; mediaObjectParam.VfeScoreboard.Value[1] = tIndex / taskParam->threadSpaceWidth; } for (aIndex = 0; aIndex < kernelParam->numArgs; aIndex++) { argParam = &kernelParam->argParams[aIndex]; index = tIndex * argParam->perThread; if ((kernelParam->cmFlags & CM_KERNEL_FLAGS_CURBE) && !argParam->perThread) { continue; } //----------------------------------------------------- CM_ASSERT(argParam->payloadOffset < kernelParam->payloadSize); //----------------------------------------------------- switch(argParam->kind) { case CM_ARGUMENT_GENERAL: MOS_SecureMemcpy( cmdInline + argParam->payloadOffset, argParam->unitSize, argParam->firstValue + index * argParam->unitSize, argParam->unitSize); break; case CM_ARGUMENT_SAMPLER: CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_SetupSamplerState( state, kernelParam, argParam, indexParam, mediaID, index, cmdInline)); break; case CM_ARGUMENT_SURFACEBUFFER: CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_SetupBufferSurfaceState( state, argParam, indexParam, bindingTable, -1, index, cmdInline)); break; case CM_ARGUMENT_SURFACE2D_UP: CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_Setup2DSurfaceUPState( state, argParam, indexParam, bindingTable, index, cmdInline)); break; case CM_ARGUMENT_SURFACE2DUP_SAMPLER: CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_Setup2DSurfaceUPSamplerState( state, argParam, indexParam, bindingTable, index, cmdInline)); break; case CM_ARGUMENT_SURFACE2D_SAMPLER: CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_Setup2DSurfaceSamplerState( state, argParam, indexParam, bindingTable, index, cmdInline)); break; case CM_ARGUMENT_SURFACE2D: CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_Setup2DSurfaceState( state, argParam, indexParam, bindingTable, index, cmdInline)); break; case CM_ARGUMENT_SURFACE3D: CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_Setup3DSurfaceState( state, argParam, indexParam, bindingTable, index, cmdInline)); break; case CM_ARGUMENT_SURFACE_VME: CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_SetupVmeSurfaceState( state, argParam, indexParam, bindingTable, 0, cmdInline)); break; case CM_ARGUMENT_SURFACE_SAMPLER8X8_AVS: case CM_ARGUMENT_SURFACE_SAMPLER8X8_VA: CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_SetupSampler8x8SurfaceState( state, argParam, indexParam, bindingTable, 0, cmdInline)); break; default: eStatus = MOS_STATUS_INVALID_PARAMETER; CM_ASSERTMESSAGE( "Argument kind '%d' is not supported", argParam->kind); goto finish; } } mediaObjectParam.pInlineData = inlineData; state->renderHal->pMhwRenderInterface->AddMediaObject(nullptr, batchBuffer, &mediaObjectParam); } } } for (i = 0; i < CM_MAX_GLOBAL_SURFACE_NUMBER; i++) { if ((kernelParam->globalSurface[i] & CM_SURFACE_MASK) != CM_NULL_SURFACE) { CM_HAL_KERNEL_ARG_PARAM tempArgParam; argParam = &tempArgParam; tempArgParam.kind = CM_ARGUMENT_SURFACEBUFFER; tempArgParam.payloadOffset = 0; tempArgParam.unitCount = 1; tempArgParam.unitSize = sizeof(uint32_t); tempArgParam.perThread = false; tempArgParam.firstValue = (uint8_t*)&kernelParam->globalSurface[i]; tempArgParam.aliasIndex = 0; tempArgParam.aliasCreated = false; CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_SetupBufferSurfaceState( state, argParam, indexParam, bindingTable, (int16_t)i, 0, nullptr)); } } // set number of samplers krnAllocation->Params.Sampler_Count = indexParam->samplerIndexCount; // add SIP surface if (kernelParam->kernelDebugEnabled) { CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_SetupSipSurfaceState(state, indexParam, bindingTable)); } finish: return eStatus; } //*----------------------------------------------------------------------------- //| Purpose: Finishes setting up HW states for the kernel //| Used by EnqueueWithHints //| Returns: Result of the operation //*----------------------------------------------------------------------------- MOS_STATUS HalCm_FinishStatesForKernelMix( PCM_HAL_STATE state, PMHW_BATCH_BUFFER batchBuffer, int32_t taskId, PCM_HAL_KERNEL_PARAM* cmExecKernels, PCM_HAL_INDEX_PARAM indexParams, int32_t *bindingTableEntries, int32_t *mediaIds, PRENDERHAL_KRN_ALLOCATION *krnAllocations, uint32_t numKernels, uint32_t hints, bool lastTask) { MOS_STATUS eStatus = MOS_STATUS_SUCCESS; PRENDERHAL_INTERFACE renderHal = state->renderHal; PMHW_MEDIA_OBJECT_PARAMS mediaObjectParams = nullptr; PCM_HAL_KERNEL_PARAM* kernelParams = nullptr; PCM_HAL_KERNEL_ARG_PARAM* argParams = nullptr; PCM_HAL_BB_ARGS bbCmArgs = nullptr; PMHW_VFE_SCOREBOARD scoreboardParams = nullptr; PCM_HAL_PARALLELISM_GRAPH_INFO parallelGraphInfo = nullptr; PCM_HAL_KERNEL_ARG_PARAM argParam = nullptr; PCM_HAL_KERNEL_SUBSLICE_INFO kernelsSliceInfo = nullptr; PCM_HAL_KERNEL_THREADSPACE_PARAM kernelTSParam = nullptr; PCM_HAL_KERNEL_GROUP_INFO groupInfo = nullptr; CM_HAL_DEPENDENCY vfeDependencyInfo ; CM_PLATFORM_INFO platformInfo ; CM_GT_SYSTEM_INFO systemInfo ; CM_HAL_SCOREBOARD_XY_MASK threadCoordinates ; uint32_t **dependRemap = nullptr; uint32_t **dispatchFreq = nullptr; uint8_t **cmdInline = nullptr; uint32_t *cmdSizes = nullptr; uint32_t *remapKrnToGrp = nullptr; uint32_t *remapGrpToKrn = nullptr; uint32_t *numKernelsPerGrp = nullptr; uint8_t *kernelScoreboardMask = nullptr; uint8_t hintsBits = 0; uint8_t tmpThreadScoreboardMask = 0; uint8_t scoreboardMask = 0; bool singleSubSlice = false; bool enableThreadSpace = false; bool kernelFound = false; bool updateCurrKernel = false; bool noDependencyCase = false; bool sufficientSliceInfo = true; uint32_t adjustedYCoord = 0; uint32_t numKernelGroups = CM_HINTS_DEFAULT_NUM_KERNEL_GRP; uint32_t totalNumThreads = 0; uint32_t hdrSize = 0; uint32_t i = 0; uint32_t j = 0; uint32_t k = 0; uint32_t tmp = 0; uint32_t tmp1 = 0; uint32_t loopCount = 0; uint32_t aIndex = 0; uint32_t index = 0; uint32_t totalReqSubSlices = 0; uint32_t difference = 0; uint32_t curKernel = 0; uint32_t numSet = 0; uint32_t numSubSlicesEnabled = 0; uint32_t sliceIndex = 0; uint32_t tmpNumSubSlice = 0; uint32_t tmpNumKernelsPerGrp = 0; uint32_t maximum = 0; uint32_t count = 0; uint32_t numDispatched = 0; uint32_t tmpIndex = 0; uint32_t numStepsDispatched = 0; uint32_t minSteps = UINT_MAX; uint32_t grpId = 0; uint32_t allocSize = 0; uint32_t currentKernel = 0; uint32_t roundRobinCount = 0; uint32_t numTasks = 0; uint32_t extraSWThreads = 0; UNUSED(taskId); CM_CHK_NULL_GOTOFINISH_MOSERROR(batchBuffer); MOS_ZeroMemory(&threadCoordinates, sizeof(CM_HAL_SCOREBOARD_XY_MASK)); MOS_ZeroMemory(&vfeDependencyInfo, sizeof(CM_HAL_DEPENDENCY)); MOS_ZeroMemory(&platformInfo, sizeof(CM_PLATFORM_INFO)); MOS_ZeroMemory(&systemInfo, sizeof(CM_GT_SYSTEM_INFO)); mediaObjectParams = (PMHW_MEDIA_OBJECT_PARAMS)MOS_AllocAndZeroMemory(sizeof(MHW_MEDIA_OBJECT_PARAMS)*numKernels); kernelParams = (PCM_HAL_KERNEL_PARAM*)MOS_AllocAndZeroMemory(sizeof(PCM_HAL_KERNEL_PARAM)*numKernels); argParams = (PCM_HAL_KERNEL_ARG_PARAM*)MOS_AllocAndZeroMemory(sizeof(PCM_HAL_KERNEL_ARG_PARAM)*numKernels); cmdInline = (uint8_t**)MOS_AllocAndZeroMemory(sizeof(uint8_t*)*numKernels); cmdSizes = (uint32_t*)MOS_AllocAndZeroMemory(sizeof(uint32_t)*numKernels); remapKrnToGrp = (uint32_t*)MOS_AllocAndZeroMemory(sizeof(uint32_t)*numKernels); remapGrpToKrn = (uint32_t*)MOS_AllocAndZeroMemory(sizeof(uint32_t)*numKernels); kernelScoreboardMask = (uint8_t*)MOS_AllocAndZeroMemory(sizeof(uint8_t)*numKernels); dependRemap = (uint32_t**)MOS_AllocAndZeroMemory(sizeof(uint32_t*)*numKernels); parallelGraphInfo = (PCM_HAL_PARALLELISM_GRAPH_INFO)MOS_AllocAndZeroMemory(sizeof(CM_HAL_PARALLELISM_GRAPH_INFO)*numKernels); dispatchFreq = (uint32_t**)MOS_AllocAndZeroMemory(sizeof(uint32_t*)*numKernels); numKernelsPerGrp = (uint32_t*)MOS_AllocAndZeroMemory(sizeof(uint32_t)*numKernels); if( !mediaObjectParams || !kernelParams || !argParams || !cmdInline || !cmdSizes || !remapKrnToGrp || !remapGrpToKrn || !kernelScoreboardMask || !dependRemap || !parallelGraphInfo || !dispatchFreq || !numKernelsPerGrp ) { eStatus = MOS_STATUS_INVALID_PARAMETER; CM_ASSERTMESSAGE("Memory allocation failed in EnqueueWithHints"); goto finish; } state->euSaturationEnabled = true; hintsBits = (hints & CM_HINTS_MASK_KERNEL_GROUPS) >> CM_HINTS_NUM_BITS_WALK_OBJ; CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_GetNumKernelsPerGroup(hintsBits, numKernels, numKernelsPerGrp, &numKernelGroups, remapKrnToGrp, remapGrpToKrn)); kernelsSliceInfo = (PCM_HAL_KERNEL_SUBSLICE_INFO)MOS_AllocAndZeroMemory(sizeof(CM_HAL_KERNEL_SUBSLICE_INFO)*numKernelGroups); groupInfo = (PCM_HAL_KERNEL_GROUP_INFO)MOS_AllocAndZeroMemory(sizeof(CM_HAL_KERNEL_GROUP_INFO)*numKernelGroups); if( !kernelsSliceInfo || !groupInfo ) { eStatus = MOS_STATUS_INVALID_PARAMETER; CM_ASSERTMESSAGE("Memory allocation failed in EnqueueWithHints"); goto finish; } for( i = 0; i < numKernelGroups; ++i) { groupInfo[i].numKernelsInGroup = numKernelsPerGrp[i]; } hdrSize = renderHal->pHwSizes->dwSizeMediaObjectHeaderCmd; for ( i = 0; i < numKernels; ++i ) { kernelParams[i] = cmExecKernels[i]; mediaObjectParams[i].dwInterfaceDescriptorOffset = mediaIds[i]; mediaObjectParams[i].dwInlineDataSize = MOS_MAX(kernelParams[i]->payloadSize, 4); cmdInline[i] = (uint8_t*)MOS_AllocAndZeroMemory(sizeof(uint8_t) * 1024); cmdSizes[i] = mediaObjectParams[i].dwInlineDataSize + hdrSize; totalNumThreads += kernelParams[i]->numThreads; } numTasks = ( hints & CM_HINTS_MASK_NUM_TASKS ) >> CM_HINTS_NUM_BITS_TASK_POS; if( numTasks > 1 ) { if( lastTask ) { extraSWThreads = totalNumThreads % numTasks; } totalNumThreads = (totalNumThreads / numTasks) + extraSWThreads; } for( i = 0; i < numKernels; ++i ) { dependRemap[i] = (uint32_t*)MOS_AllocAndZeroMemory(sizeof(uint32_t) * CM_HAL_MAX_DEPENDENCY_COUNT); for( k = 0; k < CM_HAL_MAX_DEPENDENCY_COUNT; ++k ) { // initialize each index to map to itself dependRemap[i][k] = k; } } for( i = 0; i < numKernels; ++i ) { kernelTSParam = &kernelParams[i]->kernelThreadSpaceParam; // calculate union dependency vector of all kernels with dependency if( kernelTSParam->dependencyInfo.count ) { if( vfeDependencyInfo.count == 0 ) { MOS_SecureMemcpy(&vfeDependencyInfo, sizeof(CM_HAL_DEPENDENCY), &kernelTSParam->dependencyInfo, sizeof(CM_HAL_DEPENDENCY)); kernelScoreboardMask[i] = ( 1 << vfeDependencyInfo.count ) - 1; } else { for( j = 0; j < kernelTSParam->dependencyInfo.count; ++j ) { for( k = 0; k < vfeDependencyInfo.count; ++k ) { if( (kernelTSParam->dependencyInfo.deltaX[j] == vfeDependencyInfo.deltaX[k]) && (kernelTSParam->dependencyInfo.deltaY[j] == vfeDependencyInfo.deltaY[k]) ) { CM_HAL_SETBIT(kernelScoreboardMask[i], k); dependRemap[i][j] = k; break; } } if ( k == vfeDependencyInfo.count ) { vfeDependencyInfo.deltaX[vfeDependencyInfo.count] = kernelTSParam->dependencyInfo.deltaX[j]; vfeDependencyInfo.deltaY[vfeDependencyInfo.count] = kernelTSParam->dependencyInfo.deltaY[j]; CM_HAL_SETBIT(kernelScoreboardMask[i], vfeDependencyInfo.count); vfeDependencyInfo.count++; dependRemap[i][j] = k; } } } } } // for num kernels if( vfeDependencyInfo.count > CM_HAL_MAX_DEPENDENCY_COUNT ) { eStatus = MOS_STATUS_INVALID_PARAMETER; CM_ASSERTMESSAGE("Union of kernel dependencies exceeds max dependency count (8)"); goto finish; } // set VFE scoreboarding information from union of kernel dependency vectors scoreboardParams = &state->scoreboardParams; scoreboardParams->ScoreboardMask = (uint8_t)vfeDependencyInfo.count; for( i = 0; i < scoreboardParams->ScoreboardMask; ++i ) { scoreboardParams->ScoreboardDelta[i].x = vfeDependencyInfo.deltaX[i]; scoreboardParams->ScoreboardDelta[i].y = vfeDependencyInfo.deltaY[i]; } if (vfeDependencyInfo.count == 0) { noDependencyCase = true; } CM_CHK_MOSSTATUS_GOTOFINISH(state->pfnGetPlatformInfo(state, &platformInfo, true)); singleSubSlice = (platformInfo.numSubSlices == 1) ? true : false; CM_CHK_MOSSTATUS_GOTOFINISH(state->pfnGetGTSystemInfo(state, &systemInfo)); if( !singleSubSlice ) { for( i = 0; i < numKernelGroups; ++i ) { tmpNumKernelsPerGrp = numKernelsPerGrp[i]; for( j = 0; j < tmpNumKernelsPerGrp; ++j ) { kernelTSParam = &kernelParams[count]->kernelThreadSpaceParam; switch( kernelTSParam->patternType ) { case CM_NONE_DEPENDENCY: maximum = kernelParams[count]->numThreads; break; case CM_WAVEFRONT: maximum = MOS_MIN(kernelTSParam->threadSpaceWidth, kernelTSParam->threadSpaceHeight); break; case CM_WAVEFRONT26: maximum = MOS_MIN( ((kernelTSParam->threadSpaceWidth + 1) >> 1), kernelTSParam->threadSpaceHeight); break; case CM_VERTICAL_WAVE: maximum = kernelTSParam->threadSpaceHeight; break; case CM_HORIZONTAL_WAVE: maximum = kernelTSParam->threadSpaceWidth; break; case CM_WAVEFRONT26Z: maximum = MOS_MIN( ((kernelTSParam->threadSpaceWidth - 1) >> 1), kernelTSParam->threadSpaceHeight); break; default: eStatus = MOS_STATUS_INVALID_PARAMETER; CM_ASSERTMESSAGE("Unsupported dependency pattern for EnqueueWithHints"); goto finish; } if( kernelTSParam->patternType != CM_WAVEFRONT26Z ) { CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_GetParallelGraphInfo(maximum, kernelParams[count]->numThreads, kernelTSParam->threadSpaceWidth, kernelTSParam->threadSpaceHeight, ¶llelGraphInfo[count], kernelTSParam->patternType, noDependencyCase)); } else { parallelGraphInfo[count].numSteps = kernelTSParam->dispatchInfo.numWaves; } if( kernelTSParam->patternType != CM_NONE_DEPENDENCY ) { dispatchFreq[count] = (uint32_t*)MOS_AllocAndZeroMemory(sizeof(uint32_t)*parallelGraphInfo[count].numSteps); if( !dispatchFreq[count] ) { eStatus = MOS_STATUS_INVALID_PARAMETER; CM_ASSERTMESSAGE("Memory allocation failed for EnqueueWithHints"); goto finish; } if( kernelTSParam->patternType != CM_WAVEFRONT26Z ) { CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_SetDispatchPattern(parallelGraphInfo[count], kernelTSParam->patternType, dispatchFreq[count])); } else { MOS_SecureMemcpy(dispatchFreq[count], sizeof(uint32_t)*parallelGraphInfo[count].numSteps, kernelTSParam->dispatchInfo.numThreadsInWave, sizeof(uint32_t)*parallelGraphInfo[count].numSteps); } } if (!noDependencyCase) { tmpNumSubSlice = (maximum / (platformInfo.numEUsPerSubSlice * platformInfo.numHWThreadsPerEU)) + 1; if (tmpNumSubSlice > platformInfo.numSubSlices) { tmpNumSubSlice = platformInfo.numSubSlices - 1; } if (tmpNumSubSlice > kernelsSliceInfo[i].numSubSlices) { kernelsSliceInfo[i].numSubSlices = tmpNumSubSlice; } } else { kernelsSliceInfo[i].numSubSlices = platformInfo.numSubSlices; } count++; } } if (!noDependencyCase) { for (i = 0; i < numKernelGroups; ++i) { totalReqSubSlices += kernelsSliceInfo[i].numSubSlices; } // adjust if requested less or more subslices than architecture has if (totalReqSubSlices < platformInfo.numSubSlices) { // want to add subslices starting from K0 difference = platformInfo.numSubSlices - totalReqSubSlices; tmp = tmp1 = 0; for (i = 0; i < difference; ++i) { tmp = tmp1 % numKernelGroups; kernelsSliceInfo[tmp].numSubSlices++; totalReqSubSlices++; tmp1++; } } else if (totalReqSubSlices > platformInfo.numSubSlices) { // want to subtract subslices starting from last kernel difference = totalReqSubSlices - platformInfo.numSubSlices; tmp = 0; tmp1 = numKernelGroups - 1; for (i = numKernelGroups - 1, j = 0; j < difference; --i, ++j) { tmp = tmp1 % numKernelGroups; kernelsSliceInfo[tmp].numSubSlices--; totalReqSubSlices--; tmp1 += numKernelGroups - 1; } } if (totalReqSubSlices != platformInfo.numSubSlices) { eStatus = MOS_STATUS_INVALID_PARAMETER; CM_ASSERTMESSAGE("Total requested sub-slices does not match platform's number of sub-slices"); goto finish; } } for(i = 0; i < numKernelGroups; ++i) { kernelsSliceInfo[i].destination = (PCM_HAL_KERNEL_SLICE_SUBSLICE)MOS_AllocAndZeroMemory(sizeof(CM_HAL_KERNEL_SLICE_SUBSLICE)*kernelsSliceInfo[i].numSubSlices); if( !kernelsSliceInfo[i].destination ) { eStatus = MOS_STATUS_INVALID_PARAMETER; CM_ASSERTMESSAGE("Memory allocation failed in EnqueueWithHints"); goto finish; } } // set slice, subslice for each kernel group if (systemInfo.isSliceInfoValid) { for (i = 0; i < systemInfo.numMaxSlicesSupported; ++i) { for (j = 0; j < (systemInfo.numMaxSubSlicesSupported / systemInfo.numMaxSlicesSupported); ++j) { if (systemInfo.sliceInfo[i].SubSliceInfo[j].Enabled && systemInfo.sliceInfo[i].Enabled) { if (curKernel < numKernelGroups) { if (kernelsSliceInfo[curKernel].numSubSlices == numSet) { curKernel++; numSet = 0; } } if (curKernel < numKernelGroups) { kernelsSliceInfo[curKernel].destination[numSet].slice = i; kernelsSliceInfo[curKernel].destination[numSet].subSlice = j; numSet++; } numSubSlicesEnabled++; } } } if (numSubSlicesEnabled != platformInfo.numSubSlices) { // not enough slice information, do not assign sub-slice destination sufficientSliceInfo = false; } } // set freq dispatch ratio for each group CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_SetKernelGrpFreqDispatch(parallelGraphInfo, groupInfo, numKernelGroups, &minSteps)); // set dispatch pattern for kernel with no dependency for( i = 0; i < numKernels; ++i ) { if( kernelParams[i]->kernelThreadSpaceParam.patternType == CM_NONE_DEPENDENCY ) { grpId = remapKrnToGrp[i]; allocSize = 0; if( groupInfo[grpId].freqDispatch == 0 ) { allocSize = minSteps; groupInfo[grpId].freqDispatch = 1; } else { allocSize = minSteps * groupInfo[grpId].freqDispatch; groupInfo[grpId].freqDispatch = groupInfo[grpId].freqDispatch * 2; } dispatchFreq[i] = (uint32_t*)MOS_AllocAndZeroMemory(sizeof(uint32_t)*allocSize); if( !dispatchFreq[i] ) { eStatus = MOS_STATUS_INVALID_PARAMETER; CM_ASSERTMESSAGE("Memory allocation failed in EnqueueWithHints"); goto finish; } CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_SetNoDependKernelDispatchPattern(kernelParams[i]->numThreads, allocSize, dispatchFreq[i])); } } } CM_CHK_NULL_GOTOFINISH_MOSERROR(batchBuffer->pPrivateData); bbCmArgs = (PCM_HAL_BB_ARGS) batchBuffer->pPrivateData; if( bbCmArgs->refCount > 1 ) { uint8_t *bBuffer = batchBuffer->pData + batchBuffer->iCurrent; updateCurrKernel = false; for( i = 0; i < totalNumThreads; ++i ) { if( !singleSubSlice ) { if( (dispatchFreq[currentKernel][state->hintIndexes.dispatchIndexes[currentKernel]] == numDispatched) || (state->hintIndexes.kernelIndexes[currentKernel] >= kernelParams[currentKernel]->numThreads) ) { numDispatched = 0; numStepsDispatched++; state->hintIndexes.dispatchIndexes[currentKernel]++; if( state->hintIndexes.kernelIndexes[currentKernel] >= kernelParams[currentKernel]->numThreads ) { updateCurrKernel = true; groupInfo[remapKrnToGrp[currentKernel]].numKernelsFinished++; if( groupInfo[remapKrnToGrp[currentKernel]].numKernelsFinished == groupInfo[remapKrnToGrp[currentKernel]].numKernelsInGroup ) { groupInfo[remapKrnToGrp[currentKernel]].groupFinished = 1; } else { remapGrpToKrn[tmpIndex]++; } } if( (groupInfo[remapKrnToGrp[currentKernel]].freqDispatch == numStepsDispatched) || updateCurrKernel ) { numStepsDispatched = 0; roundRobinCount++; tmpIndex = roundRobinCount % numKernelGroups; if( groupInfo[tmpIndex].groupFinished ) { loopCount = 0; while( (loopCount < numKernelGroups) && (!kernelFound) ) { roundRobinCount++; tmpIndex = roundRobinCount % numKernelGroups; if( state->hintIndexes.kernelIndexes[remapGrpToKrn[tmpIndex]] < kernelParams[remapGrpToKrn[tmpIndex]]->numThreads ) { kernelFound = true; } loopCount++; } if( !kernelFound ) { // Error shouldn't be here // if still in for loop totalNumThreads, needs to be a kernel with threads left eStatus = MOS_STATUS_UNKNOWN; CM_ASSERTMESSAGE("Couldn't find kernel with threads left for EnqueueWithHints"); goto finish; } } currentKernel = remapGrpToKrn[tmpIndex]; } } } else { if( state->hintIndexes.kernelIndexes[currentKernel] >= kernelParams[currentKernel]->numThreads ) { currentKernel++; } } if( kernelParams[currentKernel]->kernelThreadSpaceParam.threadCoordinates ) { threadCoordinates.y = kernelParams[currentKernel]->kernelThreadSpaceParam.threadCoordinates[state->hintIndexes.kernelIndexes[currentKernel]].y; threadCoordinates.mask = kernelParams[currentKernel]->kernelThreadSpaceParam.threadCoordinates[state->hintIndexes.kernelIndexes[currentKernel]].mask; enableThreadSpace = true; threadCoordinates.resetMask = kernelParams[currentKernel]->kernelThreadSpaceParam.threadCoordinates[state->hintIndexes.kernelIndexes[currentKernel]].resetMask; } if( enableThreadSpace ) { if( threadCoordinates.mask != CM_DEFAULT_THREAD_DEPENDENCY_MASK ) { tmpThreadScoreboardMask = kernelScoreboardMask[currentKernel]; // do the remapping for( k = 0; k < kernelParams[currentKernel]->kernelThreadSpaceParam.dependencyInfo.count; ++k ) { if( (threadCoordinates.mask & CM_HINTS_LEASTBIT_MASK) == 0 ) { CM_HAL_UNSETBIT(tmpThreadScoreboardMask, dependRemap[currentKernel][k]); } threadCoordinates.mask = threadCoordinates.mask >> 1; } scoreboardMask = tmpThreadScoreboardMask; } else { scoreboardMask = kernelScoreboardMask[currentKernel]; } } else { threadCoordinates.y = state->hintIndexes.kernelIndexes[currentKernel] / kernelParams[currentKernel]->kernelThreadSpaceParam.threadSpaceWidth; scoreboardMask = kernelScoreboardMask[currentKernel]; } adjustedYCoord = 0; if( currentKernel > 0 ) { // if not first kernel, and has dependency, // and along scoreboard border top need to mask out dependencies with y < 0 if( kernelScoreboardMask[currentKernel] ) { if( threadCoordinates.y == 0 ) { for( k = 0; k < vfeDependencyInfo.count; ++k ) { if( vfeDependencyInfo.deltaY[k] < 0 ) { CM_HAL_UNSETBIT(scoreboardMask, k); } } } } } if( currentKernel < numKernels - 1 ) { // if not last kernel, and has dependency, // along scoreboard border bottom need to mask out dependencies with y > 0 if( kernelScoreboardMask[currentKernel] ) { if( threadCoordinates.y == (kernelParams[currentKernel]->kernelThreadSpaceParam.threadSpaceHeight - 1)) { for( k = 0; k < vfeDependencyInfo.count; ++k) { if( vfeDependencyInfo.deltaY[k] > 0 ) { CM_HAL_UNSETBIT(scoreboardMask, k); } } } } } for( aIndex = 0; aIndex < kernelParams[currentKernel]->numArgs; aIndex++ ) { argParams[currentKernel] = &kernelParams[currentKernel]->argParams[aIndex]; index = state->hintIndexes.kernelIndexes[currentKernel] * argParams[currentKernel]->perThread; if( (kernelParams[currentKernel]->cmFlags & CM_KERNEL_FLAGS_CURBE) && !argParams[currentKernel]->perThread ) { continue; } CM_ASSERT(argParams[currentKernel]->payloadOffset < kernelParams[currentKernel]->payloadSize); switch(argParams[currentKernel]->kind) { case CM_ARGUMENT_GENERAL: break; case CM_ARGUMENT_SAMPLER: CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_SetupSamplerState( state, kernelParams[currentKernel], argParams[currentKernel], &indexParams[currentKernel], mediaIds[currentKernel], index, nullptr)); break; case CM_ARGUMENT_SURFACEBUFFER: CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_SetupBufferSurfaceState( state, argParams[currentKernel], &indexParams[currentKernel], bindingTableEntries[currentKernel], -1, index, nullptr)); break; case CM_ARGUMENT_SURFACE2D_UP: CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_Setup2DSurfaceUPState( state, argParams[currentKernel], &indexParams[currentKernel], bindingTableEntries[currentKernel], index, nullptr)); break; case CM_ARGUMENT_SURFACE2DUP_SAMPLER: CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_Setup2DSurfaceUPSamplerState( state, argParams[currentKernel], &indexParams[currentKernel], bindingTableEntries[currentKernel], index, nullptr)); break; case CM_ARGUMENT_SURFACE2D_SAMPLER: CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_Setup2DSurfaceSamplerState( state, argParams[currentKernel], &indexParams[currentKernel], bindingTableEntries[currentKernel], 0, nullptr)); break; case CM_ARGUMENT_SURFACE2D: CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_Setup2DSurfaceState( state, argParams[currentKernel], &indexParams[currentKernel], bindingTableEntries[currentKernel], index, nullptr)); break; case CM_ARGUMENT_SURFACE3D: CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_Setup3DSurfaceState( state, argParams[currentKernel], &indexParams[currentKernel], bindingTableEntries[currentKernel], index, nullptr)); break; case CM_ARGUMENT_SURFACE_VME: CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_SetupVmeSurfaceState( state, argParams[currentKernel], &indexParams[currentKernel], bindingTableEntries[currentKernel], 0, nullptr)); break; case CM_ARGUMENT_SURFACE_SAMPLER8X8_VA: case CM_ARGUMENT_SURFACE_SAMPLER8X8_AVS: CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_SetupSampler8x8SurfaceState( state, argParams[currentKernel], &indexParams[currentKernel], bindingTableEntries[currentKernel], 0, nullptr)); break; default: eStatus = MOS_STATUS_INVALID_PARAMETER; CM_ASSERTMESSAGE( "Argument kind '%d' is not supported", argParams[currentKernel]->kind); goto finish; } // switch argKind } // for numArgs if( threadCoordinates.resetMask == CM_RESET_DEPENDENCY_MASK ) { MOS_SecureMemcpy(bBuffer + (CM_SCOREBOARD_MASK_POS_IN_MEDIA_OBJECT_CMD*sizeof(uint32_t)), sizeof(uint8_t), &scoreboardMask, sizeof(uint8_t)); } batchBuffer->iCurrent += cmdSizes[currentKernel]; bBuffer += cmdSizes[currentKernel]; state->hintIndexes.kernelIndexes[currentKernel]++; enableThreadSpace = false; kernelFound = false; updateCurrKernel = false; numDispatched++; } // for totalNumThreads } // if uiRefCount > 1 else { uint8_t *bBuffer = batchBuffer->pData + batchBuffer->iCurrent; updateCurrKernel = false; for( i = 0; i < totalNumThreads; ++i) { if( !singleSubSlice ) { if( (dispatchFreq[currentKernel][state->hintIndexes.dispatchIndexes[currentKernel]] == numDispatched) || (state->hintIndexes.kernelIndexes[currentKernel] >= kernelParams[currentKernel]->numThreads) ) { numDispatched = 0; numStepsDispatched++; state->hintIndexes.dispatchIndexes[currentKernel]++; if( state->hintIndexes.kernelIndexes[currentKernel] >= kernelParams[currentKernel]->numThreads ) { updateCurrKernel = true; groupInfo[remapKrnToGrp[currentKernel]].numKernelsFinished++; if( groupInfo[remapKrnToGrp[currentKernel]].numKernelsFinished == groupInfo[remapKrnToGrp[currentKernel]].numKernelsInGroup ) { groupInfo[remapKrnToGrp[currentKernel]].groupFinished = 1; } else { remapGrpToKrn[tmpIndex]++; } } if( (groupInfo[remapKrnToGrp[currentKernel]].freqDispatch == numStepsDispatched) || updateCurrKernel ) { numStepsDispatched = 0; roundRobinCount++; tmpIndex = roundRobinCount % numKernelGroups; if( groupInfo[tmpIndex].groupFinished ) { loopCount = 0; while( (loopCount < numKernelGroups) && (!kernelFound) ) { roundRobinCount++; tmpIndex = roundRobinCount % numKernelGroups; if( state->hintIndexes.kernelIndexes[remapGrpToKrn[tmpIndex]] < kernelParams[remapGrpToKrn[tmpIndex]]->numThreads ) { kernelFound = true; } loopCount++; } if( !kernelFound ) { // Error shouldn't be here // if still in for loop totalNumThreads, needs to be a kernel with threads left eStatus = MOS_STATUS_UNKNOWN; CM_ASSERTMESSAGE("Couldn't find kernel with threads left for EnqueueWithHints"); goto finish; } } currentKernel = remapGrpToKrn[tmpIndex]; } } } else { if( state->hintIndexes.kernelIndexes[currentKernel] >= kernelParams[currentKernel]->numThreads ) { currentKernel++; } } if( kernelParams[currentKernel]->kernelThreadSpaceParam.threadCoordinates ) { threadCoordinates.x = kernelParams[currentKernel]->kernelThreadSpaceParam.threadCoordinates[state->hintIndexes.kernelIndexes[currentKernel]].x; threadCoordinates.y = kernelParams[currentKernel]->kernelThreadSpaceParam.threadCoordinates[state->hintIndexes.kernelIndexes[currentKernel]].y; threadCoordinates.mask = kernelParams[currentKernel]->kernelThreadSpaceParam.threadCoordinates[state->hintIndexes.kernelIndexes[currentKernel]].mask; enableThreadSpace = true; } mediaObjectParams[currentKernel].VfeScoreboard.ScoreboardEnable = (kernelParams[currentKernel]->kernelThreadSpaceParam.dependencyInfo.count == 0) ? 0:1; if( !singleSubSlice && systemInfo.isSliceInfoValid && sufficientSliceInfo ) { sliceIndex = kernelsSliceInfo[remapKrnToGrp[currentKernel]].counter % kernelsSliceInfo[remapKrnToGrp[currentKernel]].numSubSlices; mediaObjectParams[currentKernel].dwSliceDestinationSelect = kernelsSliceInfo[remapKrnToGrp[currentKernel]].destination[sliceIndex].slice; mediaObjectParams[currentKernel].dwHalfSliceDestinationSelect = kernelsSliceInfo[remapKrnToGrp[currentKernel]].destination[sliceIndex].subSlice; mediaObjectParams[currentKernel].bForceDestination = true; kernelsSliceInfo[remapKrnToGrp[currentKernel]].counter++; } if( enableThreadSpace ) { mediaObjectParams[currentKernel].VfeScoreboard.Value[0] = threadCoordinates.x; mediaObjectParams[currentKernel].VfeScoreboard.Value[1] = threadCoordinates.y; if( threadCoordinates.mask != CM_DEFAULT_THREAD_DEPENDENCY_MASK ) { tmpThreadScoreboardMask = kernelScoreboardMask[currentKernel]; // do the remapping for( k = 0; k < kernelParams[currentKernel]->kernelThreadSpaceParam.dependencyInfo.count; ++k ) { if( (threadCoordinates.mask & CM_HINTS_LEASTBIT_MASK) == 0 ) { CM_HAL_UNSETBIT(tmpThreadScoreboardMask, dependRemap[currentKernel][k]); } threadCoordinates.mask = threadCoordinates.mask >> 1; } mediaObjectParams[currentKernel].VfeScoreboard.ScoreboardMask = tmpThreadScoreboardMask; } else { mediaObjectParams[currentKernel].VfeScoreboard.ScoreboardMask = kernelScoreboardMask[currentKernel]; } } else { mediaObjectParams[currentKernel].VfeScoreboard.Value[0] = state->hintIndexes.kernelIndexes[currentKernel] % kernelParams[currentKernel]->kernelThreadSpaceParam.threadSpaceWidth; mediaObjectParams[currentKernel].VfeScoreboard.Value[1] = state->hintIndexes.kernelIndexes[currentKernel] / kernelParams[currentKernel]->kernelThreadSpaceParam.threadSpaceWidth; mediaObjectParams[currentKernel].VfeScoreboard.ScoreboardMask = kernelScoreboardMask[currentKernel]; } adjustedYCoord = 0; // adjust y coordinate for kernels after the first one if( currentKernel > 0 ) { // if not first kernel, and has dependency, // and along scoreboard border need to mask out dependencies with y < 0 if( kernelScoreboardMask[currentKernel] ) { if (mediaObjectParams[currentKernel].VfeScoreboard.Value[1] == 0) { for( k = 0; k < vfeDependencyInfo.count; ++k ) { if( vfeDependencyInfo.deltaY[k] < 0 ) { CM_HAL_UNSETBIT(mediaObjectParams[currentKernel].VfeScoreboard.ScoreboardMask, k); } } } } for( j = currentKernel; j > 0; --j ) { adjustedYCoord += kernelParams[j-1]->kernelThreadSpaceParam.threadSpaceHeight; } } if( currentKernel < numKernels - 1 ) { // if not last kernel, and has dependency, // along scoreboard border bottom need to mask out dependencies with y > 0 if( kernelScoreboardMask[currentKernel] ) { if (mediaObjectParams[currentKernel].VfeScoreboard.Value[1] == (kernelParams[currentKernel]->kernelThreadSpaceParam.threadSpaceHeight - 1)) { for( k = 0; k < vfeDependencyInfo.count; ++k ) { if( vfeDependencyInfo.deltaY[k] > 0 ) { CM_HAL_UNSETBIT(mediaObjectParams[currentKernel].VfeScoreboard.ScoreboardMask, k); } } } } } mediaObjectParams[currentKernel].VfeScoreboard.Value[1] = mediaObjectParams[currentKernel].VfeScoreboard.Value[1] + adjustedYCoord; for( aIndex = 0; aIndex < kernelParams[currentKernel]->numArgs; aIndex++ ) { argParams[currentKernel] = &kernelParams[currentKernel]->argParams[aIndex]; index = state->hintIndexes.kernelIndexes[currentKernel] * argParams[currentKernel]->perThread; if( (kernelParams[currentKernel]->cmFlags & CM_KERNEL_FLAGS_CURBE) && !argParams[currentKernel]->perThread ) { continue; } CM_ASSERT(argParams[currentKernel]->payloadOffset < kernelParams[currentKernel]->payloadSize); switch(argParams[currentKernel]->kind) { case CM_ARGUMENT_GENERAL: MOS_SecureMemcpy( cmdInline[currentKernel] + argParams[currentKernel]->payloadOffset, argParams[currentKernel]->unitSize, argParams[currentKernel]->firstValue + index * argParams[currentKernel]->unitSize, argParams[currentKernel]->unitSize); break; case CM_ARGUMENT_SAMPLER: CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_SetupSamplerState( state, kernelParams[currentKernel], argParams[currentKernel], &indexParams[currentKernel], mediaIds[currentKernel], index, cmdInline[currentKernel])); break; case CM_ARGUMENT_SURFACEBUFFER: CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_SetupBufferSurfaceState( state, argParams[currentKernel], &indexParams[currentKernel], bindingTableEntries[currentKernel], -1, index, cmdInline[currentKernel])); break; case CM_ARGUMENT_SURFACE2D_UP: CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_Setup2DSurfaceUPState( state, argParams[currentKernel], &indexParams[currentKernel], bindingTableEntries[currentKernel], index, cmdInline[currentKernel])); break; case CM_ARGUMENT_SURFACE2DUP_SAMPLER: CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_Setup2DSurfaceUPSamplerState( state, argParams[currentKernel], &indexParams[currentKernel], bindingTableEntries[currentKernel], index, cmdInline[currentKernel])); break; case CM_ARGUMENT_SURFACE2D_SAMPLER: CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_Setup2DSurfaceSamplerState( state, argParams[currentKernel], &indexParams[currentKernel], bindingTableEntries[currentKernel], index, cmdInline[currentKernel])); break; case CM_ARGUMENT_SURFACE2D: CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_Setup2DSurfaceState( state, argParams[currentKernel], &indexParams[currentKernel], bindingTableEntries[currentKernel], index, cmdInline[currentKernel])); break; case CM_ARGUMENT_SURFACE3D: CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_Setup3DSurfaceState( state, argParams[currentKernel], &indexParams[currentKernel], bindingTableEntries[currentKernel], index, cmdInline[currentKernel])); break; case CM_ARGUMENT_SURFACE_VME: CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_SetupVmeSurfaceState( state, argParams[currentKernel], &indexParams[currentKernel], bindingTableEntries[currentKernel], 0, cmdInline[currentKernel])); break; case CM_ARGUMENT_SURFACE_SAMPLER8X8_VA: case CM_ARGUMENT_SURFACE_SAMPLER8X8_AVS: CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_SetupSampler8x8SurfaceState( state, argParams[currentKernel], &indexParams[currentKernel], bindingTableEntries[currentKernel], 0, cmdInline[currentKernel])); break; default: eStatus = MOS_STATUS_INVALID_PARAMETER; CM_ASSERTMESSAGE( "Argument kind '%d' is not supported", argParams[currentKernel]->kind); goto finish; } } mediaObjectParams[currentKernel].pInlineData = cmdInline[currentKernel]; state->renderHal->pMhwRenderInterface->AddMediaObject(nullptr, batchBuffer, &mediaObjectParams[currentKernel]); state->hintIndexes.kernelIndexes[currentKernel]++; enableThreadSpace = false; kernelFound = false; updateCurrKernel = false; numDispatched++; } // for totalNumThreads } // else refCount <= 1 // setup global surfaces for( j = 0; j < numKernels; ++j ) { for( i = 0; i < CM_MAX_GLOBAL_SURFACE_NUMBER; ++i ) { if(( kernelParams[j]->globalSurface[i] & CM_SURFACE_MASK) != CM_NULL_SURFACE) { CM_HAL_KERNEL_ARG_PARAM tmpArgParam; argParam = &tmpArgParam; tmpArgParam.kind = CM_ARGUMENT_SURFACEBUFFER; tmpArgParam.payloadOffset = 0; tmpArgParam.unitCount = 1; tmpArgParam.unitSize = sizeof(uint32_t); tmpArgParam.perThread = false; tmpArgParam.firstValue = (uint8_t*)&kernelParams[j]->globalSurface[i]; CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_SetupBufferSurfaceState( state, argParam, &indexParams[j], bindingTableEntries[j], (int16_t)i, 0, nullptr)); } } // set number of samplers krnAllocations[j]->Params.Sampler_Count = indexParams[j].samplerIndexCount; } // check to make sure we did all threads for all kernels if (numTasks <= 1 || lastTask ) { for( i = 0; i < numKernels; ++i ) { if( state->hintIndexes.kernelIndexes[i] < kernelParams[i]->numThreads ) { eStatus = MOS_STATUS_INVALID_PARAMETER; CM_ASSERTMESSAGE("Not all threads for all kernels were put into batch buffer"); goto finish; } } } if ( lastTask ) { MOS_ZeroMemory(&state->hintIndexes.kernelIndexes, sizeof(uint32_t) * CM_MAX_TASKS_EU_SATURATION); MOS_ZeroMemory(&state->hintIndexes.dispatchIndexes, sizeof(uint32_t) * CM_MAX_TASKS_EU_SATURATION); } finish: // free memory if( mediaObjectParams ) MOS_FreeMemory(mediaObjectParams); if( kernelParams ) MOS_FreeMemory(kernelParams); if( argParams ) MOS_FreeMemory(argParams); if( cmdSizes ) MOS_FreeMemory(cmdSizes); if( remapKrnToGrp ) MOS_FreeMemory(remapKrnToGrp); if( remapGrpToKrn ) MOS_FreeMemory(remapGrpToKrn); if( kernelScoreboardMask ) MOS_FreeMemory(kernelScoreboardMask); if( parallelGraphInfo ) MOS_FreeMemory(parallelGraphInfo); if( numKernelsPerGrp ) MOS_FreeMemory(numKernelsPerGrp); if( groupInfo ) MOS_FreeMemory(groupInfo); if( cmdInline ) { for( i = 0; i < numKernels; ++i ) { if( cmdInline[i] ) MOS_FreeMemory(cmdInline[i]); } MOS_FreeMemory(cmdInline); } if( kernelsSliceInfo ) { for( i = 0; i < numKernelGroups; ++i ) { if( kernelsSliceInfo[i].destination ) MOS_FreeMemory(kernelsSliceInfo[i].destination); } MOS_FreeMemory(kernelsSliceInfo); } if( dependRemap ) { for( i = 0; i < numKernels; ++i ) { if( dependRemap[i] ) MOS_FreeMemory(dependRemap[i]); } MOS_FreeMemory(dependRemap); } if( dispatchFreq ) { for( i = 0; i < numKernels; ++i ) { if( dispatchFreq[i] ) MOS_FreeMemory(dispatchFreq[i]); } MOS_FreeMemory(dispatchFreq); } return eStatus; } uint32_t HalCm_ThreadsNumberPerGroup_MW(PCM_HAL_WALKER_PARAMS walkerParams) { int localInnerCount = 0, localMidCount = 0, localOuterCount = 0, globalInnerCount = 0, globalOuterCount = 0; int localInnerCountMax = 0, localMidCountMax = 0, localOuterCountMax = 0, globalInnerCountMax = 0; int midX = 0, midY = 0, midStep = 0; int outerX = 0, outerY = 0; int localInnerX = 0, localInnerY = 0; int blockSizeX = 0, blockSizeY = 0; //int x, y; int localLoopExecCount = walkerParams->localLoopExecCount; int globalLoopExecCount = walkerParams->globalLoopExecCount; int globalresX = walkerParams->globalResolution.x, globalresY = walkerParams->globalResolution.y; int globalOuterX = walkerParams->globalStart.x, globalOuterY = walkerParams->globalStart.y; int globalOuterStepX = walkerParams->globalOutlerLoopStride.x, globalOuterStepY = walkerParams->globalOutlerLoopStride.y; int globalInnerStepX = walkerParams->globalInnerLoopUnit.x, globalInnerStepY = walkerParams->globalInnerLoopUnit.y; int middleStepX = walkerParams->midLoopUnitX, middleStepY = walkerParams->midLoopUnitY, extraSteps = walkerParams->middleLoopExtraSteps; int localblockresX = walkerParams->blockResolution.x, localblockresY = walkerParams->blockResolution.y; int localStartX = walkerParams->localStart.x, localStartY = walkerParams->localStart.y; int localOuterStepX = walkerParams->localOutLoopStride.x, localOuterStepY = walkerParams->localOutLoopStride.y; int localInnerStepX = walkerParams->localInnerLoopUnit.x, localInnerStepY = walkerParams->localInnerLoopUnit.y; uint32_t threadsNumberPergroup = 0; //do global_outer_looper initialization while (((globalOuterX >= globalresX) && (globalInnerStepX < 0)) || (((globalOuterX + localblockresX) < 0) && (globalInnerStepX > 0)) || ((globalOuterY >= globalresY) && (globalInnerStepY < 0)) || (((globalOuterX + localblockresY) < 0) && (globalInnerStepY > 0))) { globalOuterX += globalInnerStepX; globalOuterY += globalInnerStepY; } //global_ouer_loop_in_bounds() while ((globalOuterX < globalresX) && (globalOuterY < globalresY) && (globalOuterX + localblockresX > 0) && (globalOuterY + localblockresY > 0) && (globalOuterCount <= globalLoopExecCount)) { int globalInnerX = globalOuterX; int globalInnerY = globalOuterY; if (globalInnerCountMax < globalInnerCount) { globalInnerCountMax = globalInnerCount; } globalInnerCount = 0; //global_inner_loop_in_bounds() while ((globalInnerX < globalresX) && (globalInnerY < globalresY) && (globalInnerX + localblockresX > 0) && (globalInnerY + localblockresY > 0)) { int globalInnerXCopy = globalInnerX; int globalInnerYCopy = globalInnerY; if (globalInnerX < 0) globalInnerXCopy = 0; if (globalInnerY < 0) globalInnerYCopy = 0; if (globalInnerX < 0) blockSizeX = localblockresX + globalInnerX; else if ((globalresX - globalInnerX) < localblockresX) blockSizeX = globalresX - globalInnerX; else blockSizeX = localblockresX; if (globalInnerY < 0) blockSizeY = localblockresY + globalInnerY; else if ((globalresY - globalInnerY) < localblockresY) blockSizeY = globalresY - globalInnerY; else blockSizeY = localblockresY; outerX = localStartX; outerY = localStartY; if (localOuterCountMax < localOuterCount) { localOuterCountMax = localOuterCount; } localOuterCount = 0; while ((outerX >= blockSizeX && localInnerStepX < 0) || (outerX < 0 && localInnerStepX > 0) || (outerY >= blockSizeY && localInnerStepY < 0) || (outerY < 0 && localInnerStepY > 0)) { outerX += localInnerStepX; outerY += localInnerStepY; } //local_outer_loop_in_bounds() while ((outerX < blockSizeX) && (outerY < blockSizeY) && (outerX >= 0) && (outerY >= 0) && (localOuterCount <= localLoopExecCount)) { midX = outerX; midY = outerY; midStep = 0; if (localMidCountMax < localMidCount) { localMidCountMax = localMidCount; } localMidCount = 0; //local_middle_steps_remaining() while ((midStep <= extraSteps) && (midX < blockSizeX) && (midY < blockSizeY) && (midX >= 0) && (midY >= 0)) { localInnerX = midX; localInnerY = midY; if (localInnerCountMax < localInnerCount) { localInnerCountMax = localInnerCount; } localInnerCount = 0; //local_inner_loop_shrinking() while ((localInnerX < blockSizeX) && (localInnerY < blockSizeY) && (localInnerX >= 0) && (localInnerY >= 0)) { //x = localInnerX + globalInnerXCopy; //y = localInnerY + globalInnerYCopy; localInnerCount ++; localInnerX += localInnerStepX; localInnerY += localInnerStepY; } localMidCount++; midStep++; midX += middleStepX; midY += middleStepY; } localOuterCount += 1; outerX += localOuterStepX; outerY += localOuterStepY; while ((outerX >= blockSizeX && localInnerStepX < 0) || (outerX <0 && localInnerStepX > 0) || (outerY >= blockSizeY && localInnerStepY < 0) || (outerY <0 && localInnerStepY > 0)) { outerX += localInnerStepX; outerY += localInnerStepY; } } globalInnerCount++; globalInnerX += globalInnerStepX; globalInnerY += globalInnerStepY; } globalOuterCount += 1; globalOuterX += globalOuterStepX; globalOuterY += globalOuterStepY; while (((globalOuterX >= globalresX) && (globalInnerStepX < 0)) || (((globalOuterX + localblockresX) < 0) && (globalInnerStepX > 0)) || ((globalOuterY >= globalresY) && (globalInnerStepY < 0)) || (((globalOuterX + localblockresY) < 0) && (globalInnerStepY > 0))) { globalOuterX += globalInnerStepX; globalOuterY += globalInnerStepY; } } switch (walkerParams->groupIdLoopSelect) { case CM_MW_GROUP_COLORLOOP: threadsNumberPergroup = walkerParams->colorCountMinusOne + 1; break; case CM_MW_GROUP_INNERLOCAL: threadsNumberPergroup = localInnerCount * (walkerParams->colorCountMinusOne + 1); break; case CM_MW_GROUP_MIDLOCAL: threadsNumberPergroup = localMidCount * localInnerCount * (walkerParams->colorCountMinusOne + 1); break; case CM_MW_GROUP_OUTERLOCAL: threadsNumberPergroup = localOuterCount * localMidCount * localInnerCount * (walkerParams->colorCountMinusOne + 1); break; case CM_MW_GROUP_INNERGLOBAL: threadsNumberPergroup = globalInnerCount * localOuterCount * localMidCount * localInnerCount * (walkerParams->colorCountMinusOne + 1); break; default: threadsNumberPergroup = globalOuterCount * globalInnerCount * localOuterCount * localMidCount * localInnerCount * (walkerParams->colorCountMinusOne + 1); break; } return threadsNumberPergroup; } MOS_STATUS HalCm_SetupMediaWalkerParams( PCM_HAL_STATE state, PCM_HAL_KERNEL_PARAM kernelParam) { MOS_STATUS eStatus = MOS_STATUS_SUCCESS; PCM_HAL_TASK_PARAM taskParam = state->taskParam; PCM_HAL_WALKER_PARAMS walkerParams = &kernelParam->walkerParams; //Using global walker enable flag walkerParams->cmWalkerEnable = state->walkerParams.CmWalkerEnable; if (walkerParams->cmWalkerEnable) { // MEDIA_WALKER CM_HAL_KERNEL_THREADSPACE_PARAM kernelThreadSpace; if (kernelParam->kernelThreadSpaceParam.threadSpaceWidth) { kernelThreadSpace.threadSpaceWidth = kernelParam->kernelThreadSpaceParam.threadSpaceWidth; kernelThreadSpace.threadSpaceHeight = kernelParam->kernelThreadSpaceParam.threadSpaceHeight; kernelThreadSpace.patternType = kernelParam->kernelThreadSpaceParam.patternType; kernelThreadSpace.walkingPattern = kernelParam->kernelThreadSpaceParam.walkingPattern; kernelThreadSpace.groupSelect = kernelParam->kernelThreadSpaceParam.groupSelect; kernelThreadSpace.colorCountMinusOne = kernelParam->kernelThreadSpaceParam.colorCountMinusOne; } else { kernelThreadSpace.threadSpaceWidth = (uint16_t)taskParam->threadSpaceWidth; kernelThreadSpace.threadSpaceHeight = (uint16_t)taskParam->threadSpaceHeight; kernelThreadSpace.patternType = taskParam->dependencyPattern; kernelThreadSpace.walkingPattern = taskParam->walkingPattern; kernelThreadSpace.groupSelect = taskParam->mediaWalkerGroupSelect; kernelThreadSpace.colorCountMinusOne = taskParam->colorCountMinusOne; } // check for valid thread space width and height here since different from media object if (kernelThreadSpace.threadSpaceWidth > state->cmHalInterface->GetMediaWalkerMaxThreadWidth()) { CM_ASSERTMESSAGE("Error: Exceeds the maximum thread space width."); eStatus = MOS_STATUS_INVALID_PARAMETER; goto finish; } if (kernelThreadSpace.threadSpaceHeight > state->cmHalInterface->GetMediaWalkerMaxThreadHeight()) { CM_ASSERTMESSAGE("Error: Exceeds the maximum thread space height."); eStatus = MOS_STATUS_INVALID_PARAMETER; goto finish; } //walkerParams->InterfaceDescriptorOffset = mediaID;// mediaObjectParam.dwInterfaceDescriptorOffset; walkerParams->inlineDataLength = MOS_ALIGN_CEIL(kernelParam->indirectDataParam.indirectDataSize, 4); walkerParams->inlineData = kernelParam->indirectDataParam.indirectData; walkerParams->colorCountMinusOne = kernelThreadSpace.colorCountMinusOne;// taskParam->ColorCountMinusOne; walkerParams->groupIdLoopSelect = (uint32_t)kernelThreadSpace.groupSelect; CM_WALKING_PATTERN walkPattern = kernelThreadSpace.walkingPattern; switch (kernelThreadSpace.patternType) { case CM_NONE_DEPENDENCY: break; case CM_HORIZONTAL_WAVE: walkPattern = CM_WALK_HORIZONTAL; break; case CM_VERTICAL_WAVE: walkPattern = CM_WALK_VERTICAL; break; case CM_WAVEFRONT: walkPattern = CM_WALK_WAVEFRONT; break; case CM_WAVEFRONT26: walkPattern = CM_WALK_WAVEFRONT26; break; case CM_WAVEFRONT26X: if (kernelThreadSpace.threadSpaceWidth > 1) { walkPattern = CM_WALK_WAVEFRONT26X; } else { walkPattern = CM_WALK_DEFAULT; } break; case CM_WAVEFRONT26ZIG: if (kernelThreadSpace.threadSpaceWidth > 2) { walkPattern = CM_WALK_WAVEFRONT26ZIG; } else { walkPattern = CM_WALK_DEFAULT; } break; default: CM_ASSERTMESSAGE("Error: Invalid walking pattern."); walkPattern = CM_WALK_DEFAULT; break; } if (taskParam->walkingParamsValid) { CM_CHK_MOSSTATUS_GOTOFINISH(state->cmHalInterface->SetMediaWalkerParams (taskParam->walkingParams, walkerParams)); if (walkPattern == CM_WALK_HORIZONTAL || walkPattern == CM_WALK_DEFAULT) { walkerParams->localEnd.x = walkerParams->blockResolution.x - 1; } else if (walkPattern == CM_WALK_VERTICAL) { walkerParams->localEnd.y = walkerParams->blockResolution.y - 1; } } else if (kernelParam->kernelThreadSpaceParam.walkingParamsValid) { CM_CHK_MOSSTATUS_GOTOFINISH(state->cmHalInterface->SetMediaWalkerParams( kernelParam->kernelThreadSpaceParam.walkingParams, walkerParams)); if (walkPattern == CM_WALK_HORIZONTAL || walkPattern == CM_WALK_DEFAULT) { walkerParams->localEnd.x = walkerParams->blockResolution.x - 1; } else if (walkPattern == CM_WALK_VERTICAL) { walkerParams->localEnd.y = walkerParams->blockResolution.y - 1; } } else { //Local loop parameters walkerParams->blockResolution.x = kernelThreadSpace.threadSpaceWidth; walkerParams->blockResolution.y = kernelThreadSpace.threadSpaceHeight; walkerParams->localStart.x = 0; walkerParams->localStart.y = 0; walkerParams->localEnd.x = 0; walkerParams->localEnd.y = 0; walkerParams->globalLoopExecCount = 1; walkerParams->midLoopUnitX = 0; walkerParams->midLoopUnitY = 0; walkerParams->middleLoopExtraSteps = 0; // account for odd Height/Width for 26x and 26Zig uint16_t adjHeight = ((kernelThreadSpace.threadSpaceHeight + 1) >> 1) << 1; uint16_t adjWidth = ((kernelThreadSpace.threadSpaceWidth + 1) >> 1) << 1; uint32_t maxThreadWidth = state->cmHalInterface->GetMediaWalkerMaxThreadWidth(); switch (walkPattern) { case CM_WALK_DEFAULT: case CM_WALK_HORIZONTAL: if (kernelThreadSpace.threadSpaceWidth == kernelParam->numThreads && kernelThreadSpace.threadSpaceHeight == 1) { walkerParams->blockResolution.x = MOS_MIN(kernelParam->numThreads, maxThreadWidth); walkerParams->blockResolution.y = 1 + kernelParam->numThreads / maxThreadWidth; } walkerParams->localLoopExecCount = walkerParams->blockResolution.y - 1; walkerParams->localOutLoopStride.x = 0; walkerParams->localOutLoopStride.y = 1; walkerParams->localInnerLoopUnit.x = 1; walkerParams->localInnerLoopUnit.y = 0; walkerParams->localEnd.x = walkerParams->blockResolution.x - 1; break; case CM_WALK_WAVEFRONT: walkerParams->localLoopExecCount = kernelThreadSpace.threadSpaceWidth + (kernelThreadSpace.threadSpaceHeight - 1) * 1 - 1; walkerParams->localOutLoopStride.x = 1; walkerParams->localOutLoopStride.y = 0; walkerParams->localInnerLoopUnit.x = 0xFFFF; // -1 in uint32_t:16 walkerParams->localInnerLoopUnit.y = 1; break; case CM_WALK_WAVEFRONT26: walkerParams->globalResolution.x = kernelThreadSpace.threadSpaceWidth; walkerParams->globalResolution.y = kernelThreadSpace.threadSpaceHeight; walkerParams->localOutLoopStride.x = 1; walkerParams->localOutLoopStride.y = 0; walkerParams->localInnerLoopUnit.x = 0xFFFE; // -2 in uint32_t:16 walkerParams->localInnerLoopUnit.y = 1; walkerParams->localLoopExecCount = kernelThreadSpace.threadSpaceWidth + (kernelThreadSpace.threadSpaceHeight - 1) * 2 - 1; //localLoopExecCount has limitation, it should be less than 2^12 while (walkerParams->localLoopExecCount >= 0xFFF) { //separate to multiple global levels if (walkerParams->blockResolution.x > (walkerParams->blockResolution.y * 2)) { walkerParams->blockResolution.x = (walkerParams->blockResolution.x+1) >> 1; walkerParams->globalLoopExecCount = (walkerParams->globalResolution.x + walkerParams->blockResolution.x - 1) / walkerParams->blockResolution.x; } else { walkerParams->blockResolution.y = (walkerParams->blockResolution.y + 1) >> 1; } walkerParams->localLoopExecCount = walkerParams->blockResolution.x + (walkerParams->blockResolution.y - 1) * 2 - 1; } walkerParams->globalOutlerLoopStride.x = walkerParams->blockResolution.x; walkerParams->globalOutlerLoopStride.y = 0; walkerParams->globalInnerLoopUnit.x = 0; walkerParams->globalInnerLoopUnit.y = walkerParams->blockResolution.y; break; case CM_WALK_WAVEFRONT26X: case CM_WALK_WAVEFRONT26XALT: walkerParams->localLoopExecCount = 0x7ff; walkerParams->globalLoopExecCount = 0; walkerParams->localOutLoopStride.x = 1; walkerParams->localOutLoopStride.y = 0; walkerParams->localInnerLoopUnit.x = 0xFFFE; // -2 in uint32_t:16 walkerParams->localInnerLoopUnit.y = 2; walkerParams->middleLoopExtraSteps = 1; walkerParams->midLoopUnitX = 0; walkerParams->midLoopUnitY = 1; break; case CM_WALK_WAVEFRONT26ZIG: walkerParams->localLoopExecCount = 1; walkerParams->globalLoopExecCount = (adjHeight / 2 - 1) * 2 + (adjWidth / 2) - 1; walkerParams->localOutLoopStride.x = 0; walkerParams->localOutLoopStride.y = 1; walkerParams->localInnerLoopUnit.x = 1; walkerParams->localInnerLoopUnit.y = 0; walkerParams->blockResolution.x = 2; walkerParams->blockResolution.y = 2; walkerParams->localEnd.x = walkerParams->blockResolution.x - 1; break; case CM_WALK_VERTICAL: walkerParams->localLoopExecCount = walkerParams->blockResolution.x - 1; walkerParams->localOutLoopStride.x = 1; walkerParams->localOutLoopStride.y = 0; walkerParams->localInnerLoopUnit.x = 0; walkerParams->localInnerLoopUnit.y = 1; walkerParams->localEnd.y = walkerParams->blockResolution.y - 1; break; case CM_WALK_WAVEFRONT45D: walkerParams->localLoopExecCount = 0x7ff; walkerParams->globalLoopExecCount = 0x7ff; walkerParams->localStart.x = kernelThreadSpace.threadSpaceWidth; walkerParams->localOutLoopStride.x = 1; walkerParams->localOutLoopStride.y = 0; walkerParams->localInnerLoopUnit.x = 0xFFFF; // -1 in uint32_t:16 walkerParams->localInnerLoopUnit.y = 1; break; case CM_WALK_WAVEFRONT45XD_2: walkerParams->localLoopExecCount = 0x7ff; walkerParams->globalLoopExecCount = 0x7ff; // Local walkerParams->localStart.x = kernelThreadSpace.threadSpaceWidth; walkerParams->localOutLoopStride.x = 1; walkerParams->localOutLoopStride.y = 0; walkerParams->localInnerLoopUnit.x = 0xFFFF; // -1 in uint32_t:16 walkerParams->localInnerLoopUnit.y = 2; // Mid walkerParams->middleLoopExtraSteps = 1; walkerParams->midLoopUnitX = 0; walkerParams->midLoopUnitY = 1; break; case CM_WALK_WAVEFRONT26D: walkerParams->localLoopExecCount = 0x7ff; walkerParams->globalLoopExecCount = 0x7ff; walkerParams->localStart.x = kernelThreadSpace.threadSpaceWidth; walkerParams->localOutLoopStride.x = 1; walkerParams->localOutLoopStride.y = 0; walkerParams->localInnerLoopUnit.x = 0xFFFE; // -2 in uint32_t:16 walkerParams->localInnerLoopUnit.y = 1; break; case CM_WALK_WAVEFRONT26XD: walkerParams->localLoopExecCount = 0x7ff; walkerParams->globalLoopExecCount = 0x7ff; // Local walkerParams->localStart.x = kernelThreadSpace.threadSpaceWidth; walkerParams->localOutLoopStride.x = 1; walkerParams->localOutLoopStride.y = 0; walkerParams->localInnerLoopUnit.x = 0xFFFE; // -2 in uint32_t:16 walkerParams->localInnerLoopUnit.y = 2; // Mid walkerParams->middleLoopExtraSteps = 1; walkerParams->midLoopUnitX = 0; walkerParams->midLoopUnitY = 1; break; default: walkerParams->localLoopExecCount = MOS_MIN(kernelParam->numThreads, 0x3FF); walkerParams->localOutLoopStride.x = 0; walkerParams->localOutLoopStride.y = 1; walkerParams->localInnerLoopUnit.x = 1; walkerParams->localInnerLoopUnit.y = 0; break; } //Global loop parameters: execution count, resolution and strides //Since no global loop, global resolution equals block resolution. walkerParams->globalStart.x = 0; walkerParams->globalStart.y = 0; walkerParams->globalOutlerLoopStride.y = 0; if (walkPattern == CM_WALK_WAVEFRONT26ZIG) { walkerParams->globalResolution.x = kernelThreadSpace.threadSpaceWidth; walkerParams->globalResolution.y = kernelThreadSpace.threadSpaceHeight; walkerParams->globalOutlerLoopStride.x = 2; walkerParams->globalInnerLoopUnit.x = 0xFFFC; walkerParams->globalInnerLoopUnit.y = 2; } else if(walkPattern != CM_WALK_WAVEFRONT26) { walkerParams->globalResolution.x = walkerParams->blockResolution.x; walkerParams->globalResolution.y = walkerParams->blockResolution.y; walkerParams->globalOutlerLoopStride.x = walkerParams->globalResolution.x; walkerParams->globalInnerLoopUnit.x = 0; walkerParams->globalInnerLoopUnit.y = walkerParams->globalResolution.y; } } //Need calculate number threads per group for media walker, the minimum value is 1 if (kernelThreadSpace.groupSelect > CM_MW_GROUP_NONE) { kernelParam->numberThreadsInGroup = HalCm_ThreadsNumberPerGroup_MW(walkerParams); } else { kernelParam->numberThreadsInGroup = 1; } } finish: return eStatus; } MOS_STATUS HalCm_AcquireSamplerStatistics(PCM_HAL_STATE state) { MOS_STATUS eStatus = MOS_STATUS_SUCCESS; uint32_t i = 0; unsigned int maxBTIindex[MAX_ELEMENT_TYPE_COUNT] = {0}; //tempoary variable, it will hold the max BTI index in each element type /* enumerate through the samplerTable for the one in use, then count and analyze */ for (i = 0; i < state->cmDeviceParam.maxSamplerTableSize; i++) { //state->CmDeviceParam.iMaxSamplerTableSize; if (state->samplerTable[i].bInUse) { uint32_t samplerIndex = state->samplerIndexTable[i]; if (samplerIndex != CM_INVALID_INDEX) { MHW_SAMPLER_ELEMENT_TYPE elementType = state->samplerTable[i].ElementType; maxBTIindex[elementType] = (maxBTIindex[elementType] > samplerIndex) ? maxBTIindex[elementType] : samplerIndex; } else state->samplerStatistics.samplerCount[state->samplerTable[i].ElementType]++; } } int tempbase=0; state->samplerStatistics.samplerIndexBase[MHW_Sampler2Elements] = (state->samplerStatistics.samplerCount[MHW_Sampler2Elements]) ? 0 : -1; tempbase = state->samplerStatistics.samplerIndexBase[MHW_Sampler2Elements]; state->samplerStatistics.samplerIndexBase[MHW_Sampler4Elements] = (state->samplerStatistics.samplerCount[MHW_Sampler4Elements]) ? ((tempbase == -1) ? 0 : INDEX_ALIGN(state->samplerStatistics.samplerCount[MHW_Sampler2Elements], 2, 4)) : tempbase; tempbase = state->samplerStatistics.samplerIndexBase[MHW_Sampler4Elements]; state->samplerStatistics.samplerIndexBase[MHW_Sampler8Elements] = (state->samplerStatistics.samplerCount[MHW_Sampler8Elements]) ? ((tempbase == -1) ? 0 : INDEX_ALIGN(state->samplerStatistics.samplerCount[MHW_Sampler4Elements], 4, 8)) : tempbase; tempbase = state->samplerStatistics.samplerIndexBase[MHW_Sampler8Elements]; state->samplerStatistics.samplerIndexBase[MHW_Sampler64Elements] = (state->samplerStatistics.samplerCount[MHW_Sampler64Elements]) ? ((tempbase == -1) ? 0 : INDEX_ALIGN(state->samplerStatistics.samplerCount[MHW_Sampler8Elements], 8, 64)) : tempbase; tempbase = state->samplerStatistics.samplerIndexBase[MHW_Sampler64Elements]; state->samplerStatistics.samplerIndexBase[MHW_Sampler128Elements] = (state->samplerStatistics.samplerCount[MHW_Sampler128Elements]) ? ((tempbase == -1) ? 0 : INDEX_ALIGN(state->samplerStatistics.samplerCount[MHW_Sampler64Elements], 64, 128)) : tempbase; /* There are Sampler BTI, next step needs to consider it during calculate the base */ for (int k = MHW_Sampler2Elements; k < MHW_Sampler128Elements; k++) { if (state->samplerStatistics.samplerIndexBase[k + 1] < maxBTIindex[k]) state->samplerStatistics.samplerIndexBase[k + 1] = maxBTIindex[k]; } return eStatus; } //*----------------------------------------------------------------------------- //| Purpose: Initial setup of HW states for the kernel //| Returns: Result of the operation //*----------------------------------------------------------------------------- MOS_STATUS HalCm_SetupStatesForKernelInitial( PCM_HAL_STATE state, PRENDERHAL_MEDIA_STATE mediaState, PMHW_BATCH_BUFFER batchBuffer, int32_t taskId, PCM_HAL_KERNEL_PARAM kernelParam, PCM_HAL_INDEX_PARAM indexParam, uint32_t kernelCurbeOffset, int32_t& bindingTable, int32_t& mediaID, PRENDERHAL_KRN_ALLOCATION &krnAllocation) { MOS_STATUS eStatus = MOS_STATUS_SUCCESS; PRENDERHAL_INTERFACE renderHal = state->renderHal; PRENDERHAL_STATE_HEAP stateHeap = renderHal->pStateHeap; PCM_INDIRECT_SURFACE_INFO indirectSurfaceInfo = kernelParam->indirectDataParam.surfaceInfo; PCM_GPGPU_WALKER_PARAMS perKernelGpGpuWalkerParames = &kernelParam->gpgpuWalkerParams; UNUSED(batchBuffer); UNUSED(taskId); MHW_MEDIA_OBJECT_PARAMS mediaObjectParam; PCM_HAL_KERNEL_ARG_PARAM argParam; uint32_t hdrSize; uint32_t index; uint32_t value; uint32_t btIndex; uint32_t surfIndex; uint32_t aIndex; uint32_t idZ; uint32_t idY; uint32_t idX; uint32_t localIdIndex; CM_SURFACE_BTI_INFO surfBTIInfo; bool vmeUsed = false; CM_PLATFORM_INFO platformInfo; PRENDERHAL_MEDIA_STATE_LEGACY mediaStateLegacy = (PRENDERHAL_MEDIA_STATE_LEGACY)mediaState; localIdIndex = kernelParam->localIdIndex; state->cmHalInterface->GetHwSurfaceBTIInfo(&surfBTIInfo); HalCm_PreSetBindingIndex(indexParam, CM_NULL_SURFACE_BINDING_INDEX, CM_NULL_SURFACE_BINDING_INDEX); HalCm_PreSetBindingIndex(indexParam, surfBTIInfo.reservedSurfaceStart, surfBTIInfo.reservedSurfaceStart + CM_MAX_GLOBAL_SURFACE_NUMBER - 1); if (kernelParam->indirectDataParam.surfaceCount) { for (index = 0; index < kernelParam->indirectDataParam.surfaceCount; index++) { value = (indirectSurfaceInfo + index)->bindingTableIndex; HalCm_PreSetBindingIndex(indexParam, value, value); } } // Get the binding table for this kernel CM_CHK_MOSSTATUS_GOTOFINISH(renderHal->pfnAssignBindingTable(renderHal, &bindingTable)); if (state->dshEnabled) { // Kernels are already pre-loaded in GSH // krnAllocation is the head of a linked list if (!krnAllocation) { CM_ASSERTMESSAGE("Error: Invalid kernel allocation."); goto finish; } } else { // Load the Kernel to GSH CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_LoadKernel( state, kernelParam, 0, krnAllocation)); } // initialize curbe buffer if (kernelParam->totalCurbeSize > 0) { // Update Curbe offset after curbe load command if (state->dshEnabled) { mediaStateLegacy->pDynamicState->Curbe.iCurrent += MOS_ALIGN_CEIL(kernelParam->totalCurbeSize, state->renderHal->dwCurbeBlockAlign); } else { mediaStateLegacy->iCurbeOffset += MOS_ALIGN_CEIL(kernelParam->totalCurbeSize, state->renderHal->dwCurbeBlockAlign); } } //Setup media walker parameters if it is CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_SetupMediaWalkerParams(state, kernelParam)); // Allocate Interface Descriptor mediaID = HalCm_AllocateMediaID( state, kernelParam, krnAllocation, bindingTable, kernelCurbeOffset); if (mediaID < 0) { eStatus = MOS_STATUS_INVALID_PARAMETER; CM_ASSERTMESSAGE("Unable to get Media ID"); goto finish; } // Setup the Media object hdrSize = renderHal->pHwSizes->dwSizeMediaObjectHeaderCmd; mediaObjectParam.dwInterfaceDescriptorOffset = mediaID; if (kernelParam->indirectDataParam.indirectDataSize) { mediaObjectParam.dwInlineDataSize = 0; } else { mediaObjectParam.dwInlineDataSize = MOS_MAX(kernelParam->payloadSize, 4); } // set surface state and binding table if (kernelParam->indirectDataParam.surfaceCount) { for (index = 0; index < kernelParam->indirectDataParam.surfaceCount; index++) { btIndex = (indirectSurfaceInfo + index)->bindingTableIndex; surfIndex = (indirectSurfaceInfo + index)->surfaceIndex; switch ((indirectSurfaceInfo + index)->kind) { case CM_ARGUMENT_SURFACEBUFFER: CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_SetupBufferSurfaceStateWithBTIndex( state, bindingTable, surfIndex, btIndex, 0)); break; case CM_ARGUMENT_SURFACE2D: CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_Setup2DSurfaceStateWithBTIndex( state, bindingTable, surfIndex, btIndex, 0)); break; case CM_ARGUMENT_SURFACE2D_UP: CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_Setup2DSurfaceUPStateWithBTIndex( state, bindingTable, surfIndex, btIndex, 0)); break; case CM_ARGUMENT_SURFACE2D_SAMPLER: CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_Setup2DSurfaceStateWithBTIndex( state, bindingTable, surfIndex, btIndex, 1)); break; case CM_ARGUMENT_SURFACE2DUP_SAMPLER: CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_Setup2DSurfaceUPStateWithBTIndex( state, bindingTable, surfIndex, btIndex, 1)); break; case CM_ARGUMENT_SURFACE_SAMPLER8X8_AVS: case CM_ARGUMENT_SURFACE_SAMPLER8X8_VA: CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_SetupSampler8x8SurfaceStateWithBTIndex( state, bindingTable, surfIndex, btIndex, 0, (CM_HAL_KERNEL_ARG_KIND)(indirectSurfaceInfo + index)->kind, 0)); break; case CM_ARGUMENT_SURFACE3D: CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_Setup3DSurfaceStateWithBTIndex( state, bindingTable, surfIndex, btIndex)); break; default: eStatus = MOS_STATUS_INVALID_PARAMETER; CM_ASSERTMESSAGE("Indirect Data surface kind is not supported"); goto finish; } } } // set sampler bti if (kernelParam->samplerBTIParam.samplerCount > 0) { for (uint32_t i = 0; i < kernelParam->samplerBTIParam.samplerCount; i++) { HalCm_SetupSamplerStateWithBTIndex(state, kernelParam, &kernelParam->samplerBTIParam.samplerInfo[0], i, mediaID); } } if ( ( kernelParam->curbeSizePerThread > 0 ) && ( kernelParam->stateBufferType == CM_STATE_BUFFER_NONE ) ) { uint8_t data[CM_MAX_THREAD_PAYLOAD_SIZE + 32]; uint8_t curbe[CM_MAX_CURBE_SIZE_PER_TASK + 32]; MOS_ZeroMemory(data, sizeof(data)); MOS_ZeroMemory(curbe, sizeof(curbe)); for (aIndex = 0; aIndex < kernelParam->numArgs; aIndex++) { argParam = &kernelParam->argParams[aIndex]; if (argParam->perThread || argParam->isNull) { continue; } switch (argParam->kind) { case CM_ARGUMENT_GENERAL: case CM_ARGUMENT_IMPLICT_GROUPSIZE: case CM_ARGUMENT_IMPLICT_LOCALSIZE: case CM_ARGUMENT_IMPLICIT_LOCALID: case CM_ARGUMENT_GENERAL_DEPVEC: HalCm_SetArgData(argParam, 0, data); break; case CM_ARGUMENT_SAMPLER: CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_SetupSamplerState( state, kernelParam, argParam, indexParam, mediaID, 0, data)); break; case CM_ARGUMENT_SURFACEBUFFER: CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_SetupBufferSurfaceState( state, argParam, indexParam, bindingTable, -1, 0, data)); break; case CM_ARGUMENT_SURFACE2D_UP: CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_Setup2DSurfaceUPState( state, argParam, indexParam, bindingTable, 0, data)); break; case CM_ARGUMENT_SURFACE2DUP_SAMPLER: CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_Setup2DSurfaceUPSamplerState( state, argParam, indexParam, bindingTable, 0, data)); break; case CM_ARGUMENT_SURFACE2D_SAMPLER: CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_Setup2DSurfaceSamplerState( state, argParam, indexParam, bindingTable, 0, data)); break; case CM_ARGUMENT_SURFACE2D: CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_Setup2DSurfaceState( state, argParam, indexParam, bindingTable, 0, data)); break; case CM_ARGUMENT_SURFACE3D: CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_Setup3DSurfaceState( state, argParam, indexParam, bindingTable, 0, data)); break; case CM_ARGUMENT_SURFACE_VME: // 3 surface indices CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_SetupVmeSurfaceState( state, argParam, indexParam, bindingTable, 0, data)); vmeUsed = true; break; case CM_ARGUMENT_SURFACE_SAMPLER8X8_AVS: // sampler 8x8 surface case CM_ARGUMENT_SURFACE_SAMPLER8X8_VA: // sampler 8x8 surface CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_SetupSampler8x8SurfaceState( state, argParam, indexParam, bindingTable, 0, data)); break; case CM_ARGUMENT_STATE_BUFFER: CM_CHK_MOSSTATUS_GOTOFINISH( HalCm_SetupStateBufferSurfaceState( state, argParam, indexParam, bindingTable, 0, data ) ); break; case CM_ARGUMENT_SURFACE: // Allow null surface break; case CM_ARGUMENT_SURFACE2D_SCOREBOARD: CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_Setup2DSurfaceState( state, argParam, indexParam, bindingTable, 0, data)); break; default: eStatus = MOS_STATUS_INVALID_PARAMETER; CM_ASSERTMESSAGE("Argument kind '%d' is not supported", argParam->kind); goto finish; } } if (perKernelGpGpuWalkerParames->gpgpuEnabled) { uint32_t offset = 0; uint32_t localIdXOffset = kernelParam->argParams[localIdIndex].payloadOffset; uint32_t localIdYOffset = localIdXOffset + 4; uint32_t localIdZOffset = localIdXOffset + 8; //totalCurbeSize aligned when parsing task int32_t crossThreadSize = kernelParam->crossThreadConstDataLen; //Cross thread constant data MOS_SecureMemcpy(curbe + offset, crossThreadSize, data, crossThreadSize); offset += crossThreadSize; //Per-thread data for (idZ = 0; idZ < perKernelGpGpuWalkerParames->threadDepth; idZ++) { for (idY = 0; idY < perKernelGpGpuWalkerParames->threadHeight; idY++) { for (idX = 0; idX < perKernelGpGpuWalkerParames->threadWidth; idX++) { *((uint32_t *)(data + localIdXOffset)) = idX; *((uint32_t *)(data + localIdYOffset)) = idY; *((uint32_t *)(data + localIdZOffset)) = idZ; MOS_SecureMemcpy(curbe + offset, kernelParam->curbeSizePerThread, data + crossThreadSize, kernelParam->curbeSizePerThread); offset += kernelParam->curbeSizePerThread; } } } // tell pfnLoadCurbeData the current curbe offset if (state->dshEnabled) { PRENDERHAL_MEDIA_STATE_LEGACY pCurMediaStateLegacy = (PRENDERHAL_MEDIA_STATE_LEGACY)stateHeap->pCurMediaState; PRENDERHAL_DYNAMIC_STATE dynamicState = pCurMediaStateLegacy->pDynamicState; dynamicState->Curbe.iCurrent -= MOS_ALIGN_CEIL(kernelParam->totalCurbeSize, state->renderHal->dwCurbeBlockAlign); kernelParam->curbeOffset = dynamicState->Curbe.iCurrent; } else { stateHeap->pCurMediaState->iCurbeOffset -= MOS_ALIGN_CEIL(kernelParam->totalCurbeSize, state->renderHal->dwCurbeBlockAlign); kernelParam->curbeOffset = stateHeap->pCurMediaState->iCurbeOffset; } // update curbe with data. renderHal->pfnLoadCurbeData(renderHal, stateHeap->pCurMediaState, curbe, kernelParam->totalCurbeSize); } else { CM_ASSERT(kernelParam->totalCurbeSize == kernelParam->curbeSizePerThread); // tell pfnLoadCurbeData the current curbe offset if (state->dshEnabled) { PRENDERHAL_MEDIA_STATE_LEGACY pCurMediaStateLegacy = (PRENDERHAL_MEDIA_STATE_LEGACY)stateHeap->pCurMediaState; PRENDERHAL_DYNAMIC_STATE dynamicState = pCurMediaStateLegacy->pDynamicState; dynamicState->Curbe.iCurrent -= MOS_ALIGN_CEIL(kernelParam->totalCurbeSize, state->renderHal->dwCurbeBlockAlign); kernelParam->curbeOffset = dynamicState->Curbe.iCurrent; } else { stateHeap->pCurMediaState->iCurbeOffset -= MOS_ALIGN_CEIL(kernelParam->totalCurbeSize, state->renderHal->dwCurbeBlockAlign); kernelParam->curbeOffset = stateHeap->pCurMediaState->iCurbeOffset; } // update curbe with data. renderHal->pfnLoadCurbeData(renderHal, stateHeap->pCurMediaState, data, kernelParam->totalCurbeSize); } if (state->cmHalInterface->IsOverridePowerOptionPerGpuContext() == false) // false means override per Batch. { if ((vmeUsed == true) && state->cmHalInterface->IsRequestShutdownSubslicesForVmeUsage()) { CM_CHK_MOSSTATUS_GOTOFINISH(state->pfnGetPlatformInfo(state, &platformInfo, true)); CM_POWER_OPTION cmPower; cmPower.nSlice = 1; cmPower.nSubSlice = platformInfo.numSubSlices / 2; cmPower.nEU = (uint16_t)platformInfo.numEUsPerSubSlice; state->pfnSetPowerOption(state, &cmPower); } } } #if MDF_CURBE_DATA_DUMP if (state->dumpCurbeData) { HalCm_DumpCurbeData(state); } #endif #if MDF_INTERFACE_DESCRIPTOR_DATA_DUMP if (state->dumpIDData) { HalCm_DumpInterfaceDescriptorData(state); } #endif finish: return eStatus; } MOS_STATUS HalCm_SetConditionalEndInfo( PCM_HAL_STATE state, PCM_HAL_CONDITIONAL_BB_END_INFO conditionalEndInfo, PMHW_MI_CONDITIONAL_BATCH_BUFFER_END_PARAMS conditionalBBEndParams, uint32_t index ) { if (index >= CM_MAX_CONDITIONAL_END_CMDS) { return MOS_STATUS_INVALID_PARAMETER; } MOS_ZeroMemory(&conditionalBBEndParams[index], sizeof(MHW_MI_CONDITIONAL_BATCH_BUFFER_END_PARAMS)); conditionalBBEndParams[index].presSemaphoreBuffer = &(state->bufferTable[conditionalEndInfo[index].bufferTableIndex].osResource); conditionalBBEndParams[index].dwValue = conditionalEndInfo[index].compareValue; conditionalBBEndParams[index].bDisableCompareMask = conditionalEndInfo[index].disableCompareMask; conditionalBBEndParams[index].dwOffset = conditionalEndInfo[index].offset; return MOS_STATUS_SUCCESS; } //========================================================= //*----------------------------------------------------------------------------- //| Purpose: Allocate Structures required for HW Rendering //| Returns: Result of the operation //*----------------------------------------------------------------------------- MOS_STATUS HalCm_Allocate( PCM_HAL_STATE state) // [in] Pointer to CM State { MOS_STATUS eStatus; PCM_HAL_DEVICE_PARAM deviceParam; PRENDERHAL_INTERFACE renderHal; PRENDERHAL_STATE_HEAP_SETTINGS stateHeapSettings; uint32_t i; MOS_NULL_RENDERING_FLAGS nullHWAccelerationEnable; RENDERHAL_SETTINGS_LEGACY renderHalSettings; uint32_t maxTasks; PMHW_BATCH_BUFFER batchBuffer = nullptr; //------------------------------------ CM_ASSERT(state); //------------------------------------ eStatus = MOS_STATUS_UNKNOWN; deviceParam = &state->cmDeviceParam; renderHal = state->renderHal; stateHeapSettings = &renderHal->StateHeapSettings; stateHeapSettings->iCurbeSize = CM_MAX_CURBE_SIZE_PER_TASK; stateHeapSettings->iMediaStateHeaps = deviceParam->maxTasks + 1; // + 1 to handle sync issues with current RenderHal impl (we can remove this once we insert sync value in 2nd level BB) stateHeapSettings->iMediaIDs = deviceParam->maxKernelsPerTask; // Number of Media IDs = Number of Kernels/Task stateHeapSettings->iKernelCount = deviceParam->maxGshKernelEntries; stateHeapSettings->iKernelBlockSize = deviceParam->maxKernelBinarySize; // The kernel occupied memory need be this block size aligned 256K for IVB/HSW stateHeapSettings->iKernelHeapSize = deviceParam->maxGshKernelEntries * CM_32K; // CM_MAX_GSH_KERNEL_ENTRIES * 32*1024; state->totalKernelSize = (int32_t*)MOS_AllocAndZeroMemory(sizeof(int32_t) * deviceParam->maxGshKernelEntries); if(!state->totalKernelSize) { CM_ASSERTMESSAGE("Could not allocate enough memory for state->totalKernelSize\n"); eStatus = MOS_STATUS_NO_SPACE; goto finish; } stateHeapSettings->iPerThreadScratchSize = deviceParam->maxPerThreadScratchSpaceSize; stateHeapSettings->iSipSize = CM_MAX_SIP_SIZE; stateHeapSettings->iBindingTables = deviceParam->maxKernelsPerTask; // Number of Binding tables = Number of Kernels/Task stateHeapSettings->iSurfacesPerBT = CM_MAX_SURFACE_STATES_PER_BT; // Allocate Max Binding Table indices per binding table stateHeapSettings->iSurfaceStates = CM_MAX_SURFACE_STATES; // Allocate Max Surfaces that can be indexed stateHeapSettings->iSamplersAVS = deviceParam->maxAvsSamplers; // Allocate Max AVS samplers // Initialize RenderHal Interface CM_CHK_MOSSTATUS_GOTOFINISH(renderHal->pfnInitialize(renderHal, nullptr)); // Initialize Vebox Interface if (state->veboxInterface) { CM_CHK_MOSSTATUS_GOTOFINISH(state->veboxInterface->CreateHeap()); } // Initialize the table only in Static Mode (DSH doesn't use this table at all) if (!state->dshEnabled) { // Init the data in kernel entries for Dynamic GSH for (int32_t kernelID = 0; kernelID < stateHeapSettings->iKernelCount; ++kernelID) { if (kernelID > 0) { state->totalKernelSize[kernelID] = 0; } else { state->totalKernelSize[kernelID] = stateHeapSettings->iKernelHeapSize; } } state->kernelNumInGsh = 1; } // Allocate BB (one for each media-state heap) state->numBatchBuffers = stateHeapSettings->iMediaStateHeaps; state->batchBuffers = (PMHW_BATCH_BUFFER)MOS_AllocAndZeroMemory( state->numBatchBuffers * sizeof(MHW_BATCH_BUFFER)); CM_CHK_NULL_GOTOFINISH_MOSERROR(state->batchBuffers); batchBuffer = state->batchBuffers; for (i = 0; i < (uint32_t)state->numBatchBuffers; i ++, batchBuffer ++) { batchBuffer->dwSyncTag = 0; batchBuffer->bMatch = false; batchBuffer->iPrivateType = RENDERHAL_BB_TYPE_CM; batchBuffer->iPrivateSize = sizeof(CM_HAL_BB_ARGS); batchBuffer->pPrivateData = (PCM_HAL_BB_ARGS)MOS_AllocAndZeroMemory(sizeof(CM_HAL_BB_ARGS)); CM_CHK_NULL_GOTOFINISH_MOSERROR(batchBuffer->pPrivateData); ((PCM_HAL_BB_ARGS)batchBuffer->pPrivateData)->refCount = 1; } // Allocate TimeStamp Buffer CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_AllocateTsResource(state)); // Allocate tracker resources CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_AllocateTrackerResource(state)); // Initialize dynamic general state heap CM_HAL_HEAP_PARAM heapParams; heapParams.behaviorGSH = HeapManager::Behavior::destructiveExtend; heapParams.initialSizeGSH = 0x0080000; heapParams.extendSizeGSH = 0x0080000; heapParams.trackerProducer = &state->renderHal->trackerProducer; CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_InitializeDynamicStateHeaps(state, &heapParams)); CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_AllocateTables(state)); // Allocate Task Param to hold max tasks state->taskParam = (PCM_HAL_TASK_PARAM)MOS_AllocAndZeroMemory(sizeof(CM_HAL_TASK_PARAM)); CM_CHK_NULL_GOTOFINISH_MOSERROR(state->taskParam); state->currentTaskEntry = 0; // Allocate Task TimeStamp to hold time stamps state->taskTimeStamp = (PCM_HAL_TASK_TIMESTAMP)MOS_AllocAndZeroMemory(sizeof(CM_HAL_TASK_TIMESTAMP)); CM_CHK_NULL_GOTOFINISH_MOSERROR(state->taskTimeStamp); // Setup Registration table entries state->surfaceRegTable.count = state->cmDeviceParam.max2DSurfaceTableSize; state->surfaceRegTable.entries = state->surf2DTable; maxTasks = state->cmDeviceParam.maxTasks; // Initialize the task status table MOS_FillMemory(state->taskStatusTable, (size_t)maxTasks, CM_INVALID_INDEX); // Init the null render flag nullHWAccelerationEnable = state->osInterface->pfnGetNullHWRenderFlags(state->osInterface); state->nullHwRenderCm = nullHWAccelerationEnable.Cm || nullHWAccelerationEnable.VPGobal; //during initialization stage to allocate sip resource and Get sip binary. if ((state->midThreadPreemptionDisabled == false) || (state->kernelDebugEnabled == true)) { CM_CHK_MOSSTATUS_GOTOFINISH(state->cmHalInterface->AllocateSIPCSRResource()); state->pfnGetSipBinary(state); } //Init flag for conditional batch buffer state->cbbEnabled = HalCm_IsCbbEnabled(state); //Turn Turbo boost on CM_CHK_MOSSTATUS_GOTOFINISH(state->pfnEnableTurboBoost(state)); CM_CHK_NULL_GOTOFINISH_MOSERROR(state->osInterface); state->tsFrequency = state->osInterface->pfnGetTsFrequency(state->osInterface); if (state->refactor) { state->advExecutor = CmExtensionCreator::CreateClass(); if (state->advExecutor == nullptr) { CM_ASSERTMESSAGE("Could not allocate enough memory for state->advExecutor\n"); eStatus = MOS_STATUS_NO_SPACE; goto finish; } state->advExecutor->Initialize(state); } else { state->advExecutor = nullptr; } eStatus = MOS_STATUS_SUCCESS; finish: return eStatus; } uint16_t HalCm_GetKernelPerfTag( PCM_HAL_STATE cmState, PCM_HAL_KERNEL_PARAM *kernelParams, uint32_t numKernels) { using namespace std; CM_ASSERT(cmState); CM_ASSERT(kernelParams); int perfTagKernelNum = numKernels - 1; if (numKernels > MAX_COMBINE_NUM_IN_PERFTAG) { perfTagKernelNum = MAX_COMBINE_NUM_IN_PERFTAG - 1; } // get a combined kernel name uint32_t len = numKernels * CM_MAX_KERNEL_NAME_SIZE_IN_BYTE; char *combinedName = MOS_NewArray(char, len); if (combinedName == nullptr) { // Not need to abort the process as this is only for pnp profiling CM_ASSERTMESSAGE("Error: Memory allocation error in getPertTag."); return 0; // return the default perftag } CmSafeMemSet(combinedName, 0, len); MOS_SecureStrcat(combinedName, len, kernelParams[0]->kernelName); for (uint32_t i = 1; i < numKernels; i++) { MOS_SecureStrcat(combinedName, len, ";"); MOS_SecureStrcat(combinedName, len, kernelParams[i]->kernelName); } // get perftag index int perfTagIndex = 0; map::iterator ite = cmState->perfTagIndexMap[perfTagKernelNum]->find(combinedName); if (ite == cmState->perfTagIndexMap[perfTagKernelNum]->end()) { if (cmState->currentPerfTagIndex[perfTagKernelNum] <= MAX_CUSTOMIZED_PERFTAG_INDEX) { cmState->perfTagIndexMap[perfTagKernelNum]->insert(pair(combinedName, cmState->currentPerfTagIndex[perfTagKernelNum])); perfTagIndex = cmState->currentPerfTagIndex[perfTagKernelNum] ++; } } else { perfTagIndex = ite->second; } perfTagIndex = (perfTagIndex &0xFF) | (perfTagKernelNum << 8); MosSafeDeleteArray(combinedName); return (uint16_t)perfTagIndex; } //*----------------------------------------------------------------------------- //| Purpose: Executes the CM Task //| Returns: Result of the operation //*----------------------------------------------------------------------------- MOS_STATUS HalCm_ExecuteTask( PCM_HAL_STATE state, // [in] Pointer to CM State PCM_HAL_EXEC_TASK_PARAM execParam) // [in] Pointer to Task Param { MOS_STATUS eStatus; PRENDERHAL_INTERFACE renderHal; PRENDERHAL_MEDIA_STATE mediaState; PMHW_BATCH_BUFFER batchBuffer; PCM_HAL_BB_ARGS bbCmArgs; PCM_HAL_KERNEL_PARAM kernelParam; int32_t taskId; int32_t remBindingTables; int32_t bindingTable; int32_t bti; int32_t mediaID; PRENDERHAL_KRN_ALLOCATION krnAllocations[CM_MAX_KERNELS_PER_TASK]; uint32_t vfeCurbeSize; uint32_t maxInlineDataSize, maxIndirectDataSize; uint32_t i; void *cmdBuffer = nullptr; PCM_HAL_TASK_PARAM taskParam = state->taskParam; uint32_t btsizePower2; PMOS_INTERFACE osInterface = nullptr; //----------------------------------- CM_ASSERT(state); CM_ASSERT(execParam); //----------------------------------- eStatus = MOS_STATUS_SUCCESS; renderHal = state->renderHal; mediaState = nullptr; batchBuffer = nullptr; if (execParam->numKernels > state->cmDeviceParam.maxKernelsPerTask) { eStatus = MOS_STATUS_INVALID_PARAMETER; CM_ASSERTMESSAGE("Number of Kernels per task exceeds maximum"); goto finish; } // Reset states before execute // (clear allocations, get GSH allocation index + any additional housekeeping) state->osInterface->pfnResetOsStates(state->osInterface); CM_CHK_MOSSTATUS_GOTOFINISH(renderHal->pfnReset(renderHal)); MOS_ZeroMemory(state->taskParam, sizeof(CM_HAL_TASK_PARAM)); MOS_FillMemory( state->bti2DIndexTable, state->cmDeviceParam.max2DSurfaceTableSize * sizeof( CM_HAL_MULTI_USE_BTI_ENTRY ), CM_INVALID_INDEX ); MOS_FillMemory( state->bti2DUPIndexTable, state->cmDeviceParam.max2DSurfaceUPTableSize * sizeof( CM_HAL_MULTI_USE_BTI_ENTRY ), CM_INVALID_INDEX ); MOS_FillMemory( state->bti3DIndexTable, state->cmDeviceParam.max3DSurfaceTableSize * sizeof( CM_HAL_MULTI_USE_BTI_ENTRY ), CM_INVALID_INDEX ); MOS_FillMemory( state->btiBufferIndexTable, state->cmDeviceParam.maxBufferTableSize * sizeof( CM_HAL_MULTI_USE_BTI_ENTRY ), CM_INVALID_INDEX ); MOS_FillMemory( state->samplerIndexTable, state->cmDeviceParam.maxSamplerTableSize, CM_INVALID_INDEX); MOS_FillMemory( state->sampler8x8IndexTable, state->cmDeviceParam.maxSampler8x8TableSize, CM_INVALID_INDEX); state->walkerParams.CmWalkerEnable = 0; vfeCurbeSize = 0; maxInlineDataSize = 0; maxIndirectDataSize = 0; // Get the Task Id CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_GetNewTaskId(state, &taskId)); // Parse the task CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_ParseTask(state, execParam)); // Reset the SSH configuration according to the property of the task renderHal->pStateHeap->iBindingTableSize = MOS_ALIGN_CEIL(taskParam->surfacePerBT * // Reconfigure the binding table size renderHal->pRenderHalPltInterface->GetBTStateCmdSize(), renderHal->StateHeapSettings.iBTAlignment); taskParam->surfacePerBT = renderHal->pStateHeap->iBindingTableSize/renderHal->pRenderHalPltInterface->GetBTStateCmdSize(); renderHal->StateHeapSettings.iBindingTables = renderHal->StateHeapSettings.iBindingTables * // Reconfigure the binding table number renderHal->StateHeapSettings.iSurfacesPerBT / taskParam->surfacePerBT; renderHal->StateHeapSettings.iSurfacesPerBT = taskParam->surfacePerBT; // Reconfigure the surface per BT if (execParam->numKernels > (uint32_t)renderHal->StateHeapSettings.iBindingTables) { eStatus = MOS_STATUS_INVALID_PARAMETER; CM_ASSERTMESSAGE("Number of Kernels per task exceeds the number can be hold by binding table"); goto finish; } if (execParam->kernelDebugEnabled && Mos_ResourceIsNull(&state->sipResource.osResource)) { HalCm_AllocateSipResource( state); // create sip resource if it does not exist } // Assign a MediaState from the MediaStateHeap // !!!! THIS MUST BE BEFORE Getting the BATCH_BUFFER !!! // since this method syncs the batch buffer and media state. if (state->dshEnabled) { if ( execParam->userDefinedMediaState != nullptr ) { // use exsiting media state as current state mediaState = static_cast< PRENDERHAL_MEDIA_STATE >( execParam->userDefinedMediaState ); // update current state to dsh renderHal->pStateHeap->pCurMediaState = mediaState; // Refresh sync tag for all media states in submitted queue state->criticalSectionDSH->Acquire(); renderHal->pfnRefreshSync( renderHal ); state->criticalSectionDSH->Release(); } else { // Obtain media state configuration - Curbe, Samplers (3d/AVS/VA), 8x8 sampler table, Media IDs, Kernel Spill area RENDERHAL_DYNAMIC_MEDIA_STATE_PARAMS params; state->criticalSectionDSH->Acquire(); HalCm_DSH_GetDynamicStateConfiguration( state, ¶ms, execParam->numKernels, execParam->kernels, execParam->kernelCurbeOffset ); // Prepare Media States to accommodate all parameters - Curbe, Samplers (3d/AVS/VA), 8x8 sampler table, Media IDs mediaState = renderHal->pfnAssignDynamicState( renderHal, ¶ms, RENDERHAL_COMPONENT_CM ); state->criticalSectionDSH->Release(); } } else { mediaState = renderHal->pfnAssignMediaState(renderHal, RENDERHAL_COMPONENT_CM); } CM_CHK_NULL_GOTOFINISH_MOSERROR(mediaState); // Assign/Reset SSH instance CM_CHK_MOSSTATUS_GOTOFINISH(renderHal->pfnAssignSshInstance(renderHal)); // Dynamic Batch Buffer allocation if (!state->walkerParams.CmWalkerEnable) { // Get the Batch buffer CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_GetBatchBuffer(state, execParam->numKernels, execParam->kernels, &batchBuffer)); CM_CHK_NULL_GOTOFINISH_MOSERROR(batchBuffer); bbCmArgs = (PCM_HAL_BB_ARGS)batchBuffer->pPrivateData; // Lock the batch buffer if ( (bbCmArgs->refCount == 1) || (state->taskParam->reuseBBUpdateMask == 1) ) { CM_CHK_MOSSTATUS_GOTOFINISH(renderHal->pfnLockBB(renderHal, batchBuffer)); } } if (state->useNewSamplerHeap == false) { HalCm_AcquireSamplerStatistics(state); } // Load all kernels in the same state heap - expand ISH if necessary BEFORE programming media states. // This is better than having to expand ISH in the middle of loading, when part of MediaIDs are // already programmed - not a problem in the old implementation where it would simply remove old // kernels out of the way. if (state->dshEnabled) { CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_DSH_LoadKernelArray(state, execParam->kernels, execParam->numKernels, krnAllocations)); } for (i = 0; i < execParam->numKernels; i++) { CM_HAL_INDEX_PARAM indexParam; MOS_ZeroMemory(&indexParam, sizeof(CM_HAL_INDEX_PARAM)); kernelParam = execParam->kernels[i]; CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_SetupStatesForKernelInitial(state, mediaState, batchBuffer, taskId, kernelParam, &indexParam, execParam->kernelCurbeOffset[i], bti, mediaID, krnAllocations[i])); CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_FinishStatesForKernel(state, mediaState, batchBuffer, taskId, kernelParam, i, &indexParam, bti, mediaID, krnAllocations[i])); vfeCurbeSize += MOS_ALIGN_CEIL(kernelParam->totalCurbeSize, state->renderHal->dwCurbeBlockAlign); if (kernelParam->payloadSize > maxInlineDataSize) { maxInlineDataSize = kernelParam->payloadSize; } if (kernelParam->indirectDataParam.indirectDataSize > maxIndirectDataSize) { maxIndirectDataSize = kernelParam->indirectDataParam.indirectDataSize; } if (execParam->conditionalEndBitmap & (uint64_t)1 << i) { CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_SetConditionalEndInfo(state, taskParam->conditionalEndInfo, taskParam->conditionalBBEndParams, i)); } } // Store the Max Payload Sizes in the Task params state->taskParam->vfeCurbeSize = vfeCurbeSize; if (maxIndirectDataSize) { state->taskParam->urbEntrySize = maxIndirectDataSize; } else { state->taskParam->urbEntrySize = maxInlineDataSize; } // We may have to send additional Binding table commands in command buffer. // This is needed because the surface offset (from the base on SSH) // calculation takes into account the max binding tables allocated in the // SSH. remBindingTables = renderHal->StateHeapSettings.iBindingTables - execParam->numKernels; if (remBindingTables > 0) { for (i = 0; i < (uint32_t)remBindingTables; i++) { CM_CHK_MOSSTATUS_GOTOFINISH(renderHal->pfnAssignBindingTable( renderHal, &bindingTable)); } } // until now, we know binding table index for debug surface // let's get system thread osInterface = state->osInterface; osInterface->pfnResetPerfBufferID(osInterface); if (osInterface->pfnIsPerfTagSet(osInterface) == false) { osInterface->pfnIncPerfFrameID(osInterface); uint16_t perfTag = HalCm_GetKernelPerfTag(state, execParam->kernels, execParam->numKernels); osInterface->pfnSetPerfTag(osInterface, perfTag); } #if (_RELEASE_INTERNAL || _DEBUG) #if defined(CM_DIRECT_GUC_SUPPORT) // Update the task ID table state->taskStatusTable[taskId] = (char)taskId; //for GuC direct submission, need to send out dummy command buffer to make sure PDP table got binded CM_CHK_MOSSTATUS_GOTOFINISH(state->cmHalInterface->SubmitDummyCommands( batchBuffer, taskId, execParam->kernels, &cmdBuffer)); /* make sure Dummy submission is done */ CM_HAL_QUERY_TASK_PARAM queryParam; queryParam.taskId = taskId; queryParam.status = CM_TASK_IN_PROGRESS; do { state->pfnQueryTask(state, &queryParam); } while (queryParam.status != CM_TASK_FINISHED); #endif #endif // Submit HW commands and states CM_CHK_MOSSTATUS_GOTOFINISH(state->cmHalInterface->SubmitCommands( batchBuffer, taskId, execParam->kernels, &cmdBuffer)); // Set the Task ID execParam->taskIdOut = taskId; // Set OS data if(cmdBuffer) { execParam->osData = cmdBuffer; } // Update the task ID table state->taskStatusTable[taskId] = (char)taskId; finish: if (state->dshEnabled) { state->criticalSectionDSH->Acquire(); if (mediaState && eStatus != MOS_STATUS_SUCCESS) { // Failed, release media state and heap resources renderHal->pfnReleaseDynamicState(renderHal, mediaState); } else { renderHal->pfnSubmitDynamicState(renderHal, mediaState); } state->criticalSectionDSH->Release(); } if (batchBuffer) // for Media Walker, batchBuffer is empty { if (batchBuffer->bLocked) { // Only happens in Error cases CM_CHK_NULL_RETURN_MOSERROR(batchBuffer->pPrivateData); if (((PCM_HAL_BB_ARGS)batchBuffer->pPrivateData)->refCount == 1) { renderHal->pfnUnlockBB(renderHal, batchBuffer); } } } return eStatus; } //*----------------------------------------------------------------------------- //| Purpose: Executes the CM Group Task //| Returns: Result of the operation //*----------------------------------------------------------------------------- MOS_STATUS HalCm_ExecuteGroupTask( PCM_HAL_STATE state, // [in] Pointer to CM State PCM_HAL_EXEC_GROUP_TASK_PARAM execGroupParam) // [in] Pointer to Task Param { MOS_STATUS eStatus = MOS_STATUS_SUCCESS; PRENDERHAL_INTERFACE renderHal = state->renderHal; CM_HAL_INDEX_PARAM indexParam; int32_t taskId; uint32_t remBindingTables; int32_t bindingTable; int32_t bti; int32_t mediaID; PRENDERHAL_MEDIA_STATE mediaState = nullptr; uint32_t i; void *cmdBuffer = nullptr; PCM_HAL_KERNEL_PARAM kernelParam = nullptr; PCM_HAL_TASK_PARAM taskParam = state->taskParam; uint32_t btsizePower2; uint32_t vfeCurbeSize = 0; PRENDERHAL_KRN_ALLOCATION krnAllocations[CM_MAX_KERNELS_PER_TASK]; PMOS_INTERFACE osInterface = nullptr; //----------------------------------- CM_ASSERT(state); CM_ASSERT(execGroupParam); //----------------------------------- MOS_ZeroMemory(state->taskParam, sizeof(CM_HAL_TASK_PARAM)); MOS_ZeroMemory(&indexParam, sizeof(CM_HAL_INDEX_PARAM)); MOS_FillMemory( state->bti2DIndexTable, state->cmDeviceParam.max2DSurfaceTableSize * sizeof( CM_HAL_MULTI_USE_BTI_ENTRY ), CM_INVALID_INDEX ); MOS_FillMemory( state->bti2DUPIndexTable, state->cmDeviceParam.max2DSurfaceUPTableSize * sizeof( CM_HAL_MULTI_USE_BTI_ENTRY ), CM_INVALID_INDEX ); MOS_FillMemory( state->bti3DIndexTable, state->cmDeviceParam.max3DSurfaceTableSize * sizeof( CM_HAL_MULTI_USE_BTI_ENTRY ), CM_INVALID_INDEX ); MOS_FillMemory( state->btiBufferIndexTable, state->cmDeviceParam.maxBufferTableSize * sizeof( CM_HAL_MULTI_USE_BTI_ENTRY ), CM_INVALID_INDEX ); MOS_FillMemory( state->samplerIndexTable, state->cmDeviceParam.maxSamplerTableSize, CM_INVALID_INDEX); MOS_FillMemory( state->sampler8x8IndexTable, state->cmDeviceParam.maxSampler8x8TableSize, CM_INVALID_INDEX); // Reset states before execute // (clear allocations, get GSH allocation index + any additional housekeeping) state->osInterface->pfnResetOsStates(state->osInterface); CM_CHK_MOSSTATUS_GOTOFINISH(renderHal->pfnReset(renderHal)); state->walkerParams.CmWalkerEnable = 0; state->taskParam->blGpGpuWalkerEnabled = true; // Get the Task Id CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_GetNewTaskId(state, &taskId)); // Parse the task CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_ParseGroupTask(state, execGroupParam)); // Reset the SSH configuration according to the property of the task renderHal->pStateHeap->iBindingTableSize = MOS_ALIGN_CEIL(taskParam->surfacePerBT * // Reconfigure the binding table size renderHal->pRenderHalPltInterface->GetBTStateCmdSize(), renderHal->StateHeapSettings.iBTAlignment); taskParam->surfacePerBT = renderHal->pStateHeap->iBindingTableSize / renderHal->pRenderHalPltInterface->GetBTStateCmdSize(); renderHal->StateHeapSettings.iBindingTables = renderHal->StateHeapSettings.iBindingTables * // Reconfigure the binding table number renderHal->StateHeapSettings.iSurfacesPerBT / taskParam->surfacePerBT; renderHal->StateHeapSettings.iSurfacesPerBT = taskParam->surfacePerBT; // Reconfigure the surface per BT if (execGroupParam->numKernels > (uint32_t)renderHal->StateHeapSettings.iBindingTables) { eStatus = MOS_STATUS_INVALID_PARAMETER; CM_ASSERTMESSAGE("Number of Kernels per task exceeds the number can be hold by binding table"); goto finish; } if (execGroupParam->kernelDebugEnabled && Mos_ResourceIsNull(&state->sipResource.osResource)) { HalCm_AllocateSipResource( state); // create sip resource if it does not exist } // Assign a MediaState from the MediaStateHeap // !!!! THIS MUST BE BEFORE Getting the BATCH_BUFFER !!! // since this method syncs the batch buffer and media state. if (state->dshEnabled) { if ( execGroupParam->userDefinedMediaState != nullptr ) { // Preload all kernels CM_CHK_MOSSTATUS_GOTOFINISH( HalCm_DSH_LoadKernelArray( state, execGroupParam->kernels, execGroupParam->numKernels, krnAllocations ) ); // use exsiting media state as current state mediaState = static_cast< PRENDERHAL_MEDIA_STATE >( execGroupParam->userDefinedMediaState ); // update current state to dsh renderHal->pStateHeap->pCurMediaState = mediaState; state->criticalSectionDSH->Acquire(); // Refresh sync tag for all media states in submitted queue renderHal->pfnRefreshSync( renderHal ); state->criticalSectionDSH->Release(); } else { // Preload all kernels CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_DSH_LoadKernelArray(state, execGroupParam->kernels, execGroupParam->numKernels, krnAllocations)); // Obtain media state configuration - Curbe, Samplers (3d/AVS/VA), 8x8 sampler table, Media IDs, Kernel Spill area RENDERHAL_DYNAMIC_MEDIA_STATE_PARAMS params; state->criticalSectionDSH->Acquire(); HalCm_DSH_GetDynamicStateConfiguration(state, ¶ms, execGroupParam->numKernels, execGroupParam->kernels, execGroupParam->kernelCurbeOffset); // Prepare Media States to accommodate all parameters mediaState = renderHal->pfnAssignDynamicState(renderHal, ¶ms, RENDERHAL_COMPONENT_CM); state->criticalSectionDSH->Release(); } } else { // Assign a MediaState from the MediaStateHeap // !!!! THIS MUST BE BEFORE Getting the BATCH_BUFFER !!! // since this method syncs the batch buffer and media state. mediaState = renderHal->pfnAssignMediaState(renderHal, RENDERHAL_COMPONENT_CM); } CM_CHK_NULL_GOTOFINISH_MOSERROR(mediaState); // Assign/Reset SSH instance CM_CHK_MOSSTATUS_GOTOFINISH(renderHal->pfnAssignSshInstance(renderHal)); if (state->useNewSamplerHeap == false) { HalCm_AcquireSamplerStatistics(state); } for (i = 0; i < execGroupParam->numKernels; i++) { CM_HAL_INDEX_PARAM indexParam; MOS_ZeroMemory(&indexParam, sizeof(CM_HAL_INDEX_PARAM)); kernelParam = execGroupParam->kernels[i]; CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_SetupStatesForKernelInitial(state, mediaState, nullptr, taskId, kernelParam, &indexParam, execGroupParam->kernelCurbeOffset[i], bti, mediaID, krnAllocations[i])); CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_FinishStatesForKernel(state, mediaState, nullptr, taskId, kernelParam, i, &indexParam, bti, mediaID, krnAllocations[i])); vfeCurbeSize += MOS_ALIGN_CEIL(kernelParam->totalCurbeSize, state->renderHal->dwCurbeBlockAlign); if (execGroupParam->conditionalEndBitmap & (uint64_t)1 << i) { CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_SetConditionalEndInfo(state, taskParam->conditionalEndInfo, taskParam->conditionalBBEndParams, i)); } } // Store the Max Payload Sizes in the Task params state->taskParam->vfeCurbeSize = vfeCurbeSize; state->taskParam->urbEntrySize = 0; // We may have to send additional Binding table commands in command buffer. // This is needed because the surface offset (from the base on SSH) // calculation takes into account the max binding tables allocated in the // SSH. remBindingTables = renderHal->StateHeapSettings.iBindingTables - execGroupParam->numKernels; if (remBindingTables > 0) { for (i = 0; i < remBindingTables; i++) { CM_CHK_MOSSTATUS_GOTOFINISH(renderHal->pfnAssignBindingTable( renderHal, &bindingTable)); } } // until now, we know binding table index for debug surface // let's get system thread if (execGroupParam->kernelDebugEnabled) { CM_CHK_MOSSTATUS_GOTOFINISH(state->pfnGetSipBinary(state)); } osInterface = state->osInterface; osInterface->pfnResetPerfBufferID(osInterface); if (osInterface->pfnIsPerfTagSet(osInterface) == false) { osInterface->pfnIncPerfFrameID(osInterface); int perfTag = HalCm_GetKernelPerfTag(state, execGroupParam->kernels, execGroupParam->numKernels); osInterface->pfnSetPerfTag(osInterface, (uint16_t)perfTag); } // Submit HW commands and states CM_CHK_MOSSTATUS_GOTOFINISH(state->cmHalInterface->SubmitCommands( nullptr, taskId, execGroupParam->kernels, &cmdBuffer)); // Set the Task ID execGroupParam->taskIdOut = taskId; // Set OS data if(cmdBuffer) { execGroupParam->osData = cmdBuffer; } // Update the task ID table state->taskStatusTable[taskId] = (char)taskId; finish: if (state->dshEnabled) { state->criticalSectionDSH->Acquire(); if (mediaState && eStatus != MOS_STATUS_SUCCESS) { // Failed, release media state and heap resources renderHal->pfnReleaseDynamicState(renderHal, mediaState); } else { renderHal->pfnSubmitDynamicState(renderHal, mediaState); } state->criticalSectionDSH->Release(); } return eStatus; } MOS_STATUS HalCm_ExecuteHintsTask( PCM_HAL_STATE state, // [in] Pointer to CM State PCM_HAL_EXEC_HINTS_TASK_PARAM execHintsParam) // [in] Pointer to Task Param { MOS_STATUS eStatus; PRENDERHAL_INTERFACE renderHal; PRENDERHAL_MEDIA_STATE mediaState; PMHW_BATCH_BUFFER batchBuffer; PCM_HAL_BB_ARGS bbCmArgs; PCM_HAL_KERNEL_PARAM kernelParam; uint32_t i; uint32_t numTasks; uint64_t origKernelIds[CM_MAX_KERNELS_PER_TASK]; int32_t taskId; int32_t remBindingTables; int32_t bindingTable; uint32_t vfeCurbeSize; uint32_t maxInlineDataSize; uint32_t maxIndirectDataSize; int32_t *bindingTableEntries; int32_t *mediaIds; PRENDERHAL_KRN_ALLOCATION *krnAllocations; PCM_HAL_INDEX_PARAM indexParams; bool useMediaObjects; void *cmdBuffer; bool splitTask; bool lastTask; PMOS_INTERFACE osInterface = nullptr; //------------------------------------ CM_ASSERT(state); CM_ASSERT(execHintsParam); //------------------------------------ eStatus = MOS_STATUS_SUCCESS; renderHal = state->renderHal; mediaState = nullptr; batchBuffer = nullptr; bindingTableEntries = nullptr; mediaIds = nullptr; krnAllocations = nullptr; indexParams = nullptr; useMediaObjects = false; cmdBuffer = nullptr; splitTask = false; lastTask = false; if (execHintsParam->numKernels > state->cmDeviceParam.maxKernelsPerTask) { eStatus = MOS_STATUS_INVALID_PARAMETER; CM_ASSERTMESSAGE("Number of Kernels per task exceeds maximum"); goto finish; } bindingTableEntries = (int*)MOS_AllocAndZeroMemory(sizeof(int)*execHintsParam->numKernels); mediaIds = (int*)MOS_AllocAndZeroMemory(sizeof(int)* execHintsParam->numKernels); krnAllocations = (PRENDERHAL_KRN_ALLOCATION *)MOS_AllocAndZeroMemory(sizeof(void *)* execHintsParam->numKernels); indexParams = (PCM_HAL_INDEX_PARAM)MOS_AllocAndZeroMemory(sizeof(CM_HAL_INDEX_PARAM)* execHintsParam->numKernels); if (!bindingTableEntries || !mediaIds || !krnAllocations || !indexParams) { eStatus = MOS_STATUS_INVALID_PARAMETER; CM_ASSERTMESSAGE("Memory allocation failed in ExecuteHints Task"); goto finish; } // check hints to see if need to split into multiple tasks numTasks = ( execHintsParam->hints & CM_HINTS_MASK_NUM_TASKS ) >> CM_HINTS_NUM_BITS_TASK_POS; if( numTasks > 1 ) { splitTask = true; } MOS_FillMemory(bindingTableEntries, sizeof(int) * execHintsParam->numKernels, CM_INVALID_INDEX); MOS_FillMemory(mediaIds, sizeof(int) * execHintsParam->numKernels, CM_INVALID_INDEX); MOS_FillMemory(krnAllocations, sizeof(void *)* execHintsParam->numKernels, 0); // Reset states before execute // (clear allocations, get GSH allocation index + any additional housekeeping) state->osInterface->pfnResetOsStates(state->osInterface); CM_CHK_MOSSTATUS_GOTOFINISH(renderHal->pfnReset(renderHal)); MOS_ZeroMemory(state->taskParam, sizeof(CM_HAL_TASK_PARAM)); MOS_FillMemory( state->bti2DIndexTable, state->cmDeviceParam.max2DSurfaceTableSize * sizeof( CM_HAL_MULTI_USE_BTI_ENTRY ), CM_INVALID_INDEX ); MOS_FillMemory( state->bti2DUPIndexTable, state->cmDeviceParam.max2DSurfaceUPTableSize * sizeof( CM_HAL_MULTI_USE_BTI_ENTRY ), CM_INVALID_INDEX ); MOS_FillMemory( state->bti3DIndexTable, state->cmDeviceParam.max3DSurfaceTableSize * sizeof( CM_HAL_MULTI_USE_BTI_ENTRY ), CM_INVALID_INDEX ); MOS_FillMemory( state->btiBufferIndexTable, state->cmDeviceParam.maxBufferTableSize * sizeof( CM_HAL_MULTI_USE_BTI_ENTRY ), CM_INVALID_INDEX ); MOS_FillMemory( state->samplerIndexTable, state->cmDeviceParam.maxSamplerTableSize, CM_INVALID_INDEX); MOS_FillMemory( state->sampler8x8IndexTable, state->cmDeviceParam.maxSampler8x8TableSize, CM_INVALID_INDEX); state->walkerParams.CmWalkerEnable = 0; vfeCurbeSize = 0; maxInlineDataSize = 0; maxIndirectDataSize = 0; MOS_ZeroMemory(&origKernelIds, CM_MAX_KERNELS_PER_TASK * sizeof(uint64_t)); // Get the Task Id CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_GetNewTaskId(state, &taskId)); // Parse the task CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_ParseHintsTask(state, execHintsParam)); // Assign a MediaState from the MediaStateHeap // !!!! THIS MUST BE BEFORE Getting the BATCH_BUFFER !!! // since this method syncs the batch buffer and media state. if (state->dshEnabled) { if ( execHintsParam->userDefinedMediaState != nullptr ) { // use exsiting media state as current state mediaState = static_cast< PRENDERHAL_MEDIA_STATE >( execHintsParam->userDefinedMediaState ); // update current state to dsh renderHal->pStateHeap->pCurMediaState = mediaState; // Refresh sync tag for all media states in submitted queue state->criticalSectionDSH->Acquire(); renderHal->pfnRefreshSync( renderHal ); state->criticalSectionDSH->Release(); } else { // Obtain media state configuration - Curbe, Samplers (3d/AVS/VA), 8x8 sampler table, Media IDs, Kernel Spill area RENDERHAL_DYNAMIC_MEDIA_STATE_PARAMS params; state->criticalSectionDSH->Acquire(); HalCm_DSH_GetDynamicStateConfiguration(state, ¶ms, execHintsParam->numKernels, execHintsParam->kernels, execHintsParam->kernelCurbeOffset); // Prepare Media States to accommodate all parameters - Curbe, Samplers (3d/AVS/VA), 8x8 sampler table, Media IDs mediaState = renderHal->pfnAssignDynamicState(renderHal, ¶ms, RENDERHAL_COMPONENT_CM); state->criticalSectionDSH->Release(); } } else { mediaState = renderHal->pfnAssignMediaState(renderHal, RENDERHAL_COMPONENT_CM); } CM_CHK_NULL_GOTOFINISH_MOSERROR(mediaState); if (state->useNewSamplerHeap == false) { HalCm_AcquireSamplerStatistics(state); } // Assign/Reset SSH instance CM_CHK_MOSSTATUS_GOTOFINISH(renderHal->pfnAssignSshInstance(renderHal)); if (!state->walkerParams.CmWalkerEnable) { if( splitTask ) { // save original kernel IDs for kernel binary re-use in GSH for( i = 0; i < execHintsParam->numKernels; ++i ) { origKernelIds[i] = execHintsParam->kernels[i]->kernelId; } // need to add tag to kernel IDs to distinguish batch buffer CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_AddKernelIDTag(execHintsParam->kernels, execHintsParam->numKernels, numTasks, execHintsParam->numTasksGenerated)); } // Get the Batch buffer CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_GetBatchBuffer(state, execHintsParam->numKernels, execHintsParam->kernels, &batchBuffer)); if( splitTask ) { // restore kernel IDs for kernel binary re-use in GSH for( i = 0; i < execHintsParam->numKernels; ++i ) { execHintsParam->kernels[i]->kernelId = origKernelIds[i]; } } // Lock the batch buffer CM_CHK_NULL_GOTOFINISH_MOSERROR(batchBuffer->pPrivateData); bbCmArgs = (PCM_HAL_BB_ARGS)batchBuffer->pPrivateData; if ( (bbCmArgs->refCount == 1) || ( state->taskParam->reuseBBUpdateMask == 1) ) { CM_CHK_MOSSTATUS_GOTOFINISH(renderHal->pfnLockBB(renderHal, batchBuffer)); } } // Load all kernels in the same state heap - expand ISH if necessary BEFORE programming media states. // This is better than having to expand ISH in the middle of loading, when part of MediaIDs are // already programmed - not a problem in the old implementation where it would simply remove old // kernels out of the way. if (state->dshEnabled) { CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_DSH_LoadKernelArray(state, execHintsParam->kernels, execHintsParam->numKernels, krnAllocations)); } // 0: media walker // 1: media object if( (execHintsParam->hints & CM_HINTS_MASK_MEDIAOBJECT) == CM_HINTS_MASK_MEDIAOBJECT ) { for (i = 0; i < execHintsParam->numKernels; ++i) { CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_SetupStatesForKernelInitial(state, mediaState, batchBuffer, taskId, execHintsParam->kernels[i], &indexParams[i], execHintsParam->kernelCurbeOffset[i], bindingTableEntries[i], mediaIds[i], krnAllocations[i])); } CM_CHK_NULL_GOTOFINISH_MOSERROR(batchBuffer); CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_FinishStatesForKernelMix(state, batchBuffer, taskId, execHintsParam->kernels, indexParams, bindingTableEntries, mediaIds, krnAllocations, execHintsParam->numKernels, execHintsParam->hints, execHintsParam->isLastTask)); for( i = 0; i < execHintsParam->numKernels; ++i) { kernelParam = execHintsParam->kernels[i]; vfeCurbeSize += MOS_ALIGN_CEIL(kernelParam->totalCurbeSize, state->renderHal->dwCurbeBlockAlign); if( kernelParam->payloadSize > maxInlineDataSize) { maxInlineDataSize = kernelParam->payloadSize; } if( kernelParam->indirectDataParam.indirectDataSize > maxIndirectDataSize ) { maxIndirectDataSize = kernelParam->indirectDataParam.indirectDataSize; } } // Store the Max Payload Sizes in the Task Param state->taskParam->vfeCurbeSize = vfeCurbeSize; if( maxIndirectDataSize) { state->taskParam->vfeCurbeSize = maxIndirectDataSize; } else { state->taskParam->urbEntrySize = maxInlineDataSize; } // We may have to send additional Binding table commands in command buffer. // This is needed because the surface offset (from the base on SSH) // calculation takes into account the max binding tables allocated in the // SSH. remBindingTables = state->cmDeviceParam.maxKernelsPerTask - execHintsParam->numKernels; if( remBindingTables > 0) { for( i = 0; i < (uint32_t)remBindingTables; ++i) { CM_CHK_MOSSTATUS_GOTOFINISH(renderHal->pfnAssignBindingTable( renderHal, &bindingTable)); } } osInterface = state->osInterface; osInterface->pfnResetPerfBufferID(osInterface); if (osInterface->pfnIsPerfTagSet(osInterface) == false) { osInterface->pfnIncPerfFrameID(osInterface); int perfTag = HalCm_GetKernelPerfTag(state, execHintsParam->kernels, execHintsParam->numKernels); osInterface->pfnSetPerfTag(osInterface, (uint16_t)perfTag); } // Submit HW commands and states CM_CHK_MOSSTATUS_GOTOFINISH(state->cmHalInterface->SubmitCommands( batchBuffer, taskId, execHintsParam->kernels, &cmdBuffer)); // Set the Task ID execHintsParam->taskIdOut = taskId; // Set OS data if( cmdBuffer ) { execHintsParam->osData = cmdBuffer; } // Update the task ID table state->taskStatusTable[taskId] = (char)taskId; } else { // use media walker // unimplemented for now CM_ASSERTMESSAGE("Error: Media walker is not supported."); eStatus = MOS_STATUS_UNKNOWN; } finish: if (state->dshEnabled) { state->criticalSectionDSH->Acquire(); if (mediaState && eStatus != MOS_STATUS_SUCCESS) { // Failed, release media state and heap resources renderHal->pfnReleaseDynamicState(renderHal, mediaState); } else { renderHal->pfnSubmitDynamicState(renderHal, mediaState); } state->criticalSectionDSH->Release(); } if (batchBuffer) // for MediaWalker, batchBuffer is empty { if (batchBuffer->bLocked) { // Only happens in Error cases if (batchBuffer->pPrivateData && ((PCM_HAL_BB_ARGS)batchBuffer->pPrivateData)->refCount == 1) { renderHal->pfnUnlockBB(renderHal, batchBuffer); } else if (batchBuffer->pPrivateData == nullptr) { eStatus = MOS_STATUS_NULL_POINTER; } } } // free memory if( bindingTableEntries ) MOS_FreeMemory(bindingTableEntries); if( mediaIds ) MOS_FreeMemory(mediaIds); if( krnAllocations ) MOS_FreeMemory(krnAllocations); if( indexParams ) MOS_FreeMemory( indexParams ); return eStatus; } //*----------------------------------------------------------------------------- //| Purpose: Send Commands to HW //| Returns: Get the HAL Max values //*----------------------------------------------------------------------------- MOS_STATUS HalCm_GetMaxValues( PCM_HAL_STATE state, // [in] Pointer to CM State PCM_HAL_MAX_VALUES maxValues) // [out] Pointer to Max values { PRENDERHAL_INTERFACE renderHal; renderHal = state->renderHal; maxValues->maxTasks = state->cmDeviceParam.maxTasks; maxValues->maxKernelsPerTask = CM_MAX_KERNELS_PER_TASK; maxValues->maxKernelBinarySize = state->cmDeviceParam.maxKernelBinarySize; maxValues->maxSpillSizePerHwThread = state->cmDeviceParam.maxPerThreadScratchSpaceSize; maxValues->maxSamplerTableSize = CM_MAX_SAMPLER_TABLE_SIZE; maxValues->maxBufferTableSize = CM_MAX_BUFFER_SURFACE_TABLE_SIZE; maxValues->max2DSurfaceTableSize = CM_MAX_2D_SURFACE_TABLE_SIZE; maxValues->max3DSurfaceTableSize = CM_MAX_3D_SURFACE_TABLE_SIZE; maxValues->maxArgsPerKernel = CM_MAX_ARGS_PER_KERNEL; maxValues->maxUserThreadsPerTask = CM_MAX_USER_THREADS; maxValues->maxUserThreadsPerTaskNoThreadArg = CM_MAX_USER_THREADS_NO_THREADARG; maxValues->maxArgByteSizePerKernel = CM_MAX_ARG_BYTE_PER_KERNEL; maxValues->maxSurfacesPerKernel = renderHal->pHwCaps->dwMaxBTIndex; maxValues->maxSamplersPerKernel = renderHal->pHwCaps->dwMaxUnormSamplers; maxValues->maxHwThreads = renderHal->pHwCaps->dwMaxThreads; return MOS_STATUS_SUCCESS; } //*----------------------------------------------------------------------------- //| Purpose: Get the HAL Max extended values //| Returns: Get the HAL Max extended values //*----------------------------------------------------------------------------- MOS_STATUS HalCm_GetMaxValuesEx( PCM_HAL_STATE state, // [in] Pointer to CM State PCM_HAL_MAX_VALUES_EX maxValuesEx) // [out] Pointer to extended Max values { MOS_STATUS eStatus = MOS_STATUS_SUCCESS; maxValuesEx->max2DUPSurfaceTableSize = CM_MAX_2D_SURFACE_UP_TABLE_SIZE; maxValuesEx->maxSampler8x8TableSize = CM_MAX_SAMPLER_8X8_TABLE_SIZE; maxValuesEx->maxCURBESizePerKernel = CM_MAX_CURBE_SIZE_PER_KERNEL; maxValuesEx->maxCURBESizePerTask = CM_MAX_CURBE_SIZE_PER_TASK; maxValuesEx->maxIndirectDataSizePerKernel = CM_MAX_INDIRECT_DATA_SIZE_PER_KERNEL; //MaxThreadWidth x MaxThreadHeight x ColorCount maxValuesEx->maxUserThreadsPerMediaWalker = \ state->cmHalInterface->GetMediaWalkerMaxThreadWidth()* \ state->cmHalInterface->GetMediaWalkerMaxThreadHeight() * \ CM_THREADSPACE_MAX_COLOR_COUNT; CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_GetMaxThreadCountPerThreadGroup( state, &maxValuesEx->maxUserThreadsPerThreadGroup ) ); finish: return eStatus; } //*----------------------------------------------------------------------------- //| Purpose: Register Sampler //| Returns: Result of the operation. //*----------------------------------------------------------------------------- MOS_STATUS HalCm_RegisterSampler( PCM_HAL_STATE state, // [in] Pointer to CM State PCM_HAL_SAMPLER_PARAM param) // [in] Pointer to Sampler Param { MOS_STATUS eStatus; PMHW_SAMPLER_STATE_PARAM entry; uint32_t i; eStatus = MOS_STATUS_SUCCESS; entry = nullptr; // Find a free slot for (i = 0; i < state->cmDeviceParam.maxSamplerTableSize; i++) { if (!state->samplerTable[i].bInUse) { entry = &state->samplerTable[i]; param->handle = (uint32_t)i; break; } } if (!entry) { eStatus = MOS_STATUS_INVALID_PARAMETER; CM_ASSERTMESSAGE("Sampler table is full"); goto finish; } entry->SamplerType = MHW_SAMPLER_TYPE_3D; if (state->useNewSamplerHeap == true) { entry->ElementType = MHW_Sampler1Element; } else { entry->ElementType = MHW_Sampler4Elements; } CM_CHK_MOSSTATUS_GOTOFINISH(state->pfnGetGfxMapFilter(param->minFilter, &entry->Unorm.MinFilter)); CM_CHK_MOSSTATUS_GOTOFINISH(state->pfnGetGfxMapFilter(param->magFilter, &entry->Unorm.MagFilter)); CM_CHK_MOSSTATUS_GOTOFINISH(state->pfnGetGfxTextAddress(param->addressU, &entry->Unorm.AddressU)); CM_CHK_MOSSTATUS_GOTOFINISH(state->pfnGetGfxTextAddress(param->addressV, &entry->Unorm.AddressV)); CM_CHK_MOSSTATUS_GOTOFINISH(state->pfnGetGfxTextAddress(param->addressW, &entry->Unorm.AddressW)); entry->Unorm.SurfaceFormat = (MHW_SAMPLER_SURFACE_PIXEL_TYPE)param->surfaceFormat; switch (entry->Unorm.SurfaceFormat) { case MHW_SAMPLER_SURFACE_PIXEL_UINT: entry->Unorm.BorderColorRedU = param->borderColorRedU; entry->Unorm.BorderColorGreenU = param->borderColorGreenU; entry->Unorm.BorderColorBlueU = param->borderColorBlueU; entry->Unorm.BorderColorAlphaU = param->borderColorAlphaU; break; case MHW_SAMPLER_SURFACE_PIXEL_SINT: entry->Unorm.BorderColorRedS = param->borderColorRedS; entry->Unorm.BorderColorGreenS = param->borderColorGreenS; entry->Unorm.BorderColorBlueS = param->borderColorBlueS; entry->Unorm.BorderColorAlphaS = param->borderColorAlphaS; break; default: entry->Unorm.BorderColorRedF = param->borderColorRedF; entry->Unorm.BorderColorGreenF = param->borderColorGreenF; entry->Unorm.BorderColorBlueF = param->borderColorBlueF; entry->Unorm.BorderColorAlphaF = param->borderColorAlphaF; } entry->Unorm.bBorderColorIsValid = true; entry->bInUse = true; finish: return eStatus; } //*----------------------------------------------------------------------------- //| Purpose: UnRegister Sampler //| Returns: Result of the operation. //*----------------------------------------------------------------------------- MOS_STATUS HalCm_UnRegisterSampler( PCM_HAL_STATE state, // [in] Pointer to CM State uint32_t handle) // [in] Pointer to Sampler Param { MOS_STATUS eStatus; PMHW_SAMPLER_STATE_PARAM entry; eStatus = MOS_STATUS_SUCCESS; if (handle >= state->cmDeviceParam.maxSamplerTableSize) { eStatus = MOS_STATUS_INVALID_HANDLE; CM_ASSERTMESSAGE("Invalid handle '%d'", handle); goto finish; } entry = &state->samplerTable[handle]; // need to clear the state entirely instead of just setting bInUse to false MOS_ZeroMemory(entry, sizeof(MHW_SAMPLER_STATE_PARAM)); finish: return eStatus; } //*----------------------------------------------------------------------------- //| Purpose: Register Sampler8x8 //| Returns: Result of the operation. //*----------------------------------------------------------------------------- MOS_STATUS HalCm_RegisterSampler8x8( PCM_HAL_STATE state, PCM_HAL_SAMPLER_8X8_PARAM param) { return state->cmHalInterface->RegisterSampler8x8(param); } //*----------------------------------------------------------------------------- //| Purpose: UnRegister Sampler //| Returns: Result of the operation. //*----------------------------------------------------------------------------- MOS_STATUS HalCm_UnRegisterSampler8x8( PCM_HAL_STATE state, // [in] Pointer to CM State uint32_t handle) // [in] Pointer to Sampler8x8 Param { MOS_STATUS eStatus; uint32_t index8x8; PMHW_SAMPLER_STATE_PARAM entry; PCM_HAL_SAMPLER_8X8_ENTRY sampler8x8Entry; eStatus = MOS_STATUS_SUCCESS; if (handle >= state->cmDeviceParam.maxSamplerTableSize) { eStatus = MOS_STATUS_INVALID_HANDLE; CM_ASSERTMESSAGE("Invalid handle '%d'", handle); goto finish; } entry = &state->samplerTable[handle]; entry->bInUse = false; if ( entry->SamplerType == MHW_SAMPLER_TYPE_AVS ) { index8x8 = entry->Avs.stateID; if ( index8x8 >= state->cmDeviceParam.maxSampler8x8TableSize ) { eStatus = MOS_STATUS_INVALID_HANDLE; CM_ASSERTMESSAGE( "Invalid 8x8 handle '%d'", handle ); goto finish; } sampler8x8Entry = &state->sampler8x8Table[ index8x8 ]; sampler8x8Entry->inUse = false; } // need to clear the state entirely instead of just setting bInUse to false MOS_ZeroMemory(entry, sizeof(MHW_SAMPLER_STATE_PARAM)); finish: return eStatus; } //*----------------------------------------------------------------------------- //| Purpose: Frees the buffer and removes from the table //| Returns: Result of the operation. //*----------------------------------------------------------------------------- MOS_STATUS HalCm_FreeBuffer( PCM_HAL_STATE state, // [in] Pointer to CM State uint32_t handle) // [in] Pointer to Buffer Param { MOS_STATUS eStatus; PCM_HAL_BUFFER_ENTRY entry; PMOS_INTERFACE osInterface; MOS_GFXRES_FREE_FLAGS resFreeFlags = {0}; resFreeFlags.AssumeNotInUse = 1; eStatus = MOS_STATUS_SUCCESS; osInterface = state->osInterface; // Get the Buffer Entry CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_GetBufferEntry(state, handle, &entry)); if (state->advExecutor) { state->advExecutor->DeleteBufferStateMgr(entry->surfStateMgr); } if (entry->isAllocatedbyCmrtUmd) { osInterface->pfnFreeResourceWithFlag(osInterface, &entry->osResource, resFreeFlags.Value); } else { HalCm_OsResource_Unreference(&entry->osResource); } osInterface->pfnResetResourceAllocationIndex(osInterface, &entry->osResource); entry->size = 0; entry->address = nullptr; finish: return eStatus; } //*----------------------------------------------------------------------------- //| Purpose: Set surface read flag used in on demand sync //| Returns: Result of the operation. //*----------------------------------------------------------------------------- MOS_STATUS HalCm_SetSurfaceReadFlag( PCM_HAL_STATE state, // [in] Pointer to CM State uint32_t handle, // [in] index of surface 2d bool readSync, MOS_GPU_CONTEXT gpuContext) { MOS_STATUS eStatus = MOS_STATUS_SUCCESS; PCM_HAL_SURFACE2D_ENTRY entry; // Get the Buffer Entry CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_GetSurface2DEntry(state, handle, &entry)); if (HalCm_IsValidGpuContext(gpuContext)) { entry->readSyncs[gpuContext] = readSync; if (state->advExecutor) { state->advExecutor->Set2DRenderTarget(entry->surfStateMgr, !readSync); } } else { return MOS_STATUS_UNKNOWN; } finish: return eStatus; } //*----------------------------------------------------------------------------- //| Purpose: Read the data from buffer and return //| Returns: Result of the operation. //*----------------------------------------------------------------------------- MOS_STATUS HalCm_LockBuffer( PCM_HAL_STATE state, // [in] Pointer to CM State PCM_HAL_BUFFER_PARAM param) // [in] Pointer to Buffer Param { MOS_STATUS eStatus; PCM_HAL_BUFFER_ENTRY entry; PMOS_INTERFACE osInterface; MOS_LOCK_PARAMS lockFlags; eStatus = MOS_STATUS_SUCCESS; osInterface = state->osInterface; CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_GetBufferEntry(state, param->handle, &entry)); if ((param->lockFlag != CM_HAL_LOCKFLAG_READONLY) && (param->lockFlag != CM_HAL_LOCKFLAG_WRITEONLY) ) { eStatus = MOS_STATUS_INVALID_HANDLE; CM_ASSERTMESSAGE("Invalid lock flag!"); eStatus = MOS_STATUS_UNKNOWN; goto finish; } // RegisterResource will be called in AddResourceToHWCmd. It is not allowed to be called by hal explicitly if (!osInterface->apoMosEnabled) { CM_CHK_HRESULT_GOTOFINISH_MOSERROR( osInterface->pfnRegisterResource(osInterface, &entry->osResource, true, true)); } // Lock the resource MOS_ZeroMemory(&lockFlags, sizeof(MOS_LOCK_PARAMS)); if (param->lockFlag == CM_HAL_LOCKFLAG_READONLY) { lockFlags.ReadOnly = true; } else { lockFlags.WriteOnly = true; } lockFlags.ForceCached = true; param->data = osInterface->pfnLockResource( osInterface, &entry->osResource, &lockFlags); CM_CHK_NULL_GOTOFINISH_MOSERROR(param->data); finish: return eStatus; } //*----------------------------------------------------------------------------- //| Purpose: Writes the data to buffer //| Returns: Result of the operation. //*----------------------------------------------------------------------------- MOS_STATUS HalCm_UnlockBuffer( PCM_HAL_STATE state, // [in] Pointer to CM State PCM_HAL_BUFFER_PARAM param) // [in] Pointer to Buffer Param { MOS_STATUS eStatus; PCM_HAL_BUFFER_ENTRY entry; PMOS_INTERFACE osInterface; eStatus = MOS_STATUS_SUCCESS; osInterface = state->osInterface; CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_GetBufferEntry(state, param->handle, &entry)); CM_CHK_HRESULT_GOTOFINISH_MOSERROR(osInterface->pfnUnlockResource(osInterface, &entry->osResource)); finish: return eStatus; } //*----------------------------------------------------------------------------- //| Purpose: Frees the buffer and removes from the table //| Returns: Result of the operation. //*----------------------------------------------------------------------------- MOS_STATUS HalCm_FreeSurface2DUP( PCM_HAL_STATE state, // [in] Pointer to CM State uint32_t handle) // [in] Pointer to Buffer Param { MOS_STATUS eStatus; PCM_HAL_SURFACE2D_UP_ENTRY entry; PMOS_INTERFACE osInterface; MOS_GFXRES_FREE_FLAGS resFreeFlags = {0}; resFreeFlags.AssumeNotInUse = 1; eStatus = MOS_STATUS_SUCCESS; osInterface = state->osInterface; // Get the Buffer Entry CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_GetResourceUPEntry(state, handle, &entry)); if (state->advExecutor) { state->advExecutor->Delete2Dor3DStateMgr(entry->surfStateMgr); } osInterface->pfnFreeResourceWithFlag(osInterface, &entry->osResource, resFreeFlags.Value); osInterface->pfnResetResourceAllocationIndex(osInterface, &entry->osResource); entry->width = 0; finish: return eStatus; } //*----------------------------------------------------------------------------- //| Purpose: Get 2D surface pitch and physical size //| Returns: Result of the operation. //*----------------------------------------------------------------------------- MOS_STATUS HalCm_GetSurface2DTileYPitch( PCM_HAL_STATE state, // [in] Pointer to CM State PCM_HAL_SURFACE2D_PARAM param) // [in] Pointer to Buffer Param { MOS_STATUS eStatus; MOS_SURFACE surface; PRENDERHAL_INTERFACE renderHal; uint32_t index; RENDERHAL_GET_SURFACE_INFO info; //----------------------------------------------- CM_ASSERT(state); //----------------------------------------------- eStatus = MOS_STATUS_UNKNOWN; renderHal = state->renderHal; index = param->handle; // Get Details of 2D surface and fill the surface MOS_ZeroMemory(&surface, sizeof(surface)); surface.OsResource = state->umdSurf2DTable[index].osResource; surface.dwWidth = state->umdSurf2DTable[index].width; surface.dwHeight = state->umdSurf2DTable[index].height; surface.Format = state->umdSurf2DTable[index].format; surface.dwDepth = 1; MOS_ZeroMemory(&info, sizeof(RENDERHAL_GET_SURFACE_INFO)); CM_CHK_MOSSTATUS_GOTOFINISH(RenderHal_GetSurfaceInfo( state->osInterface, &info, &surface)); param->pitch = surface.dwPitch; finish: return eStatus; } //*----------------------------------------------------------------------------- //| Purpose: Sets width and height values for 2D surface state //| Returns: Result of the operation. //*----------------------------------------------------------------------------- MOS_STATUS HalCm_Set2DSurfaceStateParam( PCM_HAL_STATE state, PCM_HAL_SURFACE2D_SURFACE_STATE_PARAM param, uint32_t aliasIndex, uint32_t handle) { MOS_STATUS eStatus = MOS_STATUS_SUCCESS; CM_CHK_NULL_GOTOFINISH_MOSERROR(state); CM_CHK_NULL_GOTOFINISH_MOSERROR(param); if (aliasIndex < state->surfaceArraySize) { state->umdSurf2DTable[handle].surfStateSet = true; } state->umdSurf2DTable[handle].surfaceStateParam[ aliasIndex / state->surfaceArraySize] = *param; finish: return eStatus; } //*----------------------------------------------------------------------------- //| Purpose: Sets width and height values for 2D surface state //| Returns: Result of the operation. //*----------------------------------------------------------------------------- MOS_STATUS HalCm_SetBufferSurfaceStateParameters( PCM_HAL_STATE state, PCM_HAL_BUFFER_SURFACE_STATE_PARAM param) { MOS_STATUS eStatus; uint32_t size; uint32_t offset; uint32_t index; uint32_t aliasIndex; CM_CHK_NULL_GOTOFINISH_MOSERROR(state); CM_CHK_NULL_GOTOFINISH_MOSERROR(param); eStatus = MOS_STATUS_SUCCESS; index = param->handle; aliasIndex = param->aliasIndex; if (aliasIndex < state->surfaceArraySize) state->bufferTable[index].surfStateSet = true; state->bufferTable[index].surfaceStateEntry[aliasIndex / state->surfaceArraySize].surfaceStateSize = param->size; state->bufferTable[index].surfaceStateEntry[aliasIndex / state->surfaceArraySize].surfaceStateOffset = param->offset; state->bufferTable[index].surfaceStateEntry[aliasIndex / state->surfaceArraySize].surfaceStateMOCS = param->mocs; finish: return eStatus; } //*----------------------------------------------------------------------------- //| Purpose: Sets mocs value for surface //| Returns: Result of the operation. //*----------------------------------------------------------------------------- MOS_STATUS HalCm_SetSurfaceMOCS( PCM_HAL_STATE state, uint32_t handle, uint16_t mocs, uint32_t argKind) { MOS_STATUS eStatus = MOS_STATUS_SUCCESS; switch (argKind) { case CM_ARGUMENT_SURFACEBUFFER: state->bufferTable[handle].memObjCtl = mocs; if (state->advExecutor) { state->advExecutor->SetBufferMemoryObjectControl(state->bufferTable[handle].surfStateMgr, mocs); } break; case CM_ARGUMENT_SURFACE2D: case CM_ARGUMENT_SURFACE2D_SAMPLER: case CM_ARGUMENT_SURFACE_SAMPLER8X8_AVS: case CM_ARGUMENT_SURFACE_SAMPLER8X8_VA: state->umdSurf2DTable[handle].memObjCtl = mocs; if (state->advExecutor) { state->advExecutor->Set2Dor3DMemoryObjectControl(state->umdSurf2DTable[handle].surfStateMgr, mocs); } break; case CM_ARGUMENT_SURFACE2D_UP: case CM_ARGUMENT_SURFACE2DUP_SAMPLER: state->surf2DUPTable[handle].memObjCtl = mocs; if (state->advExecutor) { state->advExecutor->Set2Dor3DMemoryObjectControl(state->surf2DUPTable[handle].surfStateMgr, mocs); } break; case CM_ARGUMENT_SURFACE3D: state->surf3DTable[handle].memObjCtl = mocs; if (state->advExecutor) { state->advExecutor->Set2Dor3DMemoryObjectControl(state->surf3DTable[handle].surfStateMgr, mocs); } break; default: eStatus = MOS_STATUS_INVALID_PARAMETER; CM_ASSERTMESSAGE("Invalid argument type in MOCS settings"); goto finish; } finish: return eStatus; } //*----------------------------------------------------------------------------- //| Purpose: Allocate surface 2D //| Returns: Result of the operation. //*----------------------------------------------------------------------------- MOS_STATUS HalCm_AllocateSurface2D( PCM_HAL_STATE state, // [in] Pointer to CM State PCM_HAL_SURFACE2D_PARAM param) // [in] Pointer to surface 2D Param { MOS_STATUS eStatus; PMOS_INTERFACE osInterface; PCM_HAL_SURFACE2D_ENTRY entry = nullptr; MOS_ALLOC_GFXRES_PARAMS allocParams; uint32_t i; //----------------------------------------------- CM_ASSERT(param->width > 0); //----------------------------------------------- eStatus = MOS_STATUS_SUCCESS; osInterface = state->osInterface; // Find a free slot for (i = 0; i < state->cmDeviceParam.max2DSurfaceTableSize; i++) { if(Mos_ResourceIsNull(&state->umdSurf2DTable[i].osResource)) { entry = &state->umdSurf2DTable[i]; param->handle = (uint32_t)i; break; } } if (!entry) { eStatus = MOS_STATUS_INVALID_PARAMETER; CM_ASSERTMESSAGE("Surface2D table is full"); goto finish; } if(param->isAllocatedbyCmrtUmd) { MOS_ZeroMemory(&allocParams, sizeof(MOS_ALLOC_GFXRES_PARAMS)); allocParams.Type = MOS_GFXRES_2D; allocParams.dwWidth = param->width; allocParams.dwHeight = param->height; allocParams.pSystemMemory = param->data; allocParams.Format = param->format; allocParams.TileType = MOS_TILE_Y; allocParams.pBufName = "CmSurface2D"; CM_CHK_HRESULT_GOTOFINISH_MOSERROR(osInterface->pfnAllocateResource( osInterface, &allocParams, &entry->osResource)); entry->width = param->width; entry->height = param->height; entry->format = param->format; entry->isAllocatedbyCmrtUmd = param->isAllocatedbyCmrtUmd; } else { entry->width = param->width; entry->height = param->height; entry->format = param->format; entry->isAllocatedbyCmrtUmd = false; entry->osResource = *param->mosResource; HalCm_OsResource_Reference(&entry->osResource); } // set default CM MOS usage entry->memObjCtl = (state->cmHalInterface->GetDefaultMOCS()) << 8; if (state->advExecutor) { entry->surfStateMgr = state->advExecutor->Create2DStateMgr(&entry->osResource); state->advExecutor->Set2Dor3DOrigFormat(entry->surfStateMgr, entry->format); state->advExecutor->Set2Dor3DOrigDimension(entry->surfStateMgr, entry->width, entry->height, 0); // no need to change depth in 2D surface } for (int i = 0; i < CM_HAL_GPU_CONTEXT_COUNT; i++) { entry->readSyncs[i] = false; } finish: return eStatus; } //*----------------------------------------------------------------------------- //| Purpose: Allocate surface 2D //| Returns: Result of the operation. //*----------------------------------------------------------------------------- MOS_STATUS HalCm_UpdateSurface2D( PCM_HAL_STATE state, // [in] Pointer to CM State PCM_HAL_SURFACE2D_PARAM param) // [in] Pointer to surface 2D Param { MOS_STATUS hr; PMOS_INTERFACE osInterface; PCM_HAL_SURFACE2D_ENTRY entry = nullptr; MOS_ALLOC_GFXRES_PARAMS allocParams; uint32_t i = param->handle; //----------------------------------------------- CM_ASSERT(param->width > 0); //----------------------------------------------- hr = MOS_STATUS_SUCCESS; osInterface = state->osInterface; entry = &state->umdSurf2DTable[i]; HalCm_OsResource_Unreference(&entry->osResource); entry->width = param->width; entry->height = param->height; entry->format = param->format; entry->isAllocatedbyCmrtUmd = false; entry->osResource = *param->mosResource; HalCm_OsResource_Reference(&entry->osResource); if (state->advExecutor) { state->advExecutor->Delete2Dor3DStateMgr(entry->surfStateMgr); entry->surfStateMgr = state->advExecutor->Create2DStateMgr(&entry->osResource); state->advExecutor->Set2Dor3DOrigFormat(entry->surfStateMgr, entry->format); state->advExecutor->Set2Dor3DOrigDimension(entry->surfStateMgr, entry->width, entry->height, 0); // no need to change depth in 2D surface } for (int i = 0; i < CM_HAL_GPU_CONTEXT_COUNT; i++) { entry->readSyncs[i] = false; } return hr; } //*----------------------------------------------------------------------------- //| Purpose: Allocate Linear Buffer or BufferUP //| Returns: Result of the operation. //*----------------------------------------------------------------------------- MOS_STATUS HalCm_UpdateBuffer( PCM_HAL_STATE state, // [in] Pointer to CM State PCM_HAL_BUFFER_PARAM param) // [in] Pointer to Buffer Param { MOS_STATUS hr; PMOS_INTERFACE osInterface; PCM_HAL_BUFFER_ENTRY entry = nullptr; MOS_ALLOC_GFXRES_PARAMS allocParams; uint32_t i = param->handle; PMOS_RESOURCE osResource; //----------------------------------------------- CM_ASSERT(param->size > 0); //----------------------------------------------- hr = MOS_STATUS_SUCCESS; osInterface = state->renderHal->pOsInterface; entry = &state->bufferTable[i]; HalCm_OsResource_Unreference(&entry->osResource); entry->osResource = *param->mosResource; HalCm_OsResource_Reference(&entry->osResource); entry->size = param->size; entry->isAllocatedbyCmrtUmd = false; entry->surfaceStateEntry[0].surfaceStateSize = entry->size; entry->surfaceStateEntry[0].surfaceStateOffset = 0; entry->surfaceStateEntry[0].surfaceStateMOCS = 0; if (state->advExecutor) { state->advExecutor->DeleteBufferStateMgr(entry->surfStateMgr); entry->surfStateMgr = state->advExecutor->CreateBufferStateMgr(&entry->osResource); state->advExecutor->SetBufferOrigSize(entry->surfStateMgr, entry->size); } return hr; } //*----------------------------------------------------------------------------- //| Purpose: Frees the surface 2D and removes from the table //| Returns: Result of the operation. //*----------------------------------------------------------------------------- MOS_STATUS HalCm_FreeSurface2D( PCM_HAL_STATE state, // [in] Pointer to CM State uint32_t handle) // [in] Pointer to Buffer Param { MOS_STATUS eStatus; PCM_HAL_SURFACE2D_ENTRY entry; PMOS_INTERFACE osInterface; MOS_GFXRES_FREE_FLAGS resFreeFlags = {0}; resFreeFlags.AssumeNotInUse = 1; eStatus = MOS_STATUS_SUCCESS; osInterface = state->osInterface; // Get the Buffer Entry CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_GetSurface2DEntry(state, handle, &entry)); if (state->advExecutor) { state->advExecutor->Delete2Dor3DStateMgr(entry->surfStateMgr); } if(entry->isAllocatedbyCmrtUmd) { osInterface->pfnFreeResourceWithFlag(osInterface, &entry->osResource, resFreeFlags.Value); } else { HalCm_OsResource_Unreference(&entry->osResource); } MOS_ZeroMemory(&entry->osResource, sizeof(entry->osResource)); entry->width = 0; entry->height = 0; entry->frameType = CM_FRAME; for (int i = 0; i < CM_HAL_GPU_CONTEXT_COUNT; i++) { entry->readSyncs[i] = false; } finish: return eStatus; } //*----------------------------------------------------------------------------- //| Purpose: Allocate 3D resource //| Returns: Result of the operation. //*----------------------------------------------------------------------------- MOS_STATUS HalCm_AllocateSurface3D(CM_HAL_STATE *state, // [in] Pointer to CM State CM_HAL_3DRESOURCE_PARAM *param) // [in] Pointer to Buffer Param) { MOS_STATUS eStatus = MOS_STATUS_SUCCESS; //----------------------------------------------- CM_ASSERT(state); CM_ASSERT(param->depth > 1); CM_ASSERT(param->width > 0); CM_ASSERT(param->height > 0); //----------------------------------------------- MOS_INTERFACE *osInterface = state->osInterface; // Finds a free slot. CM_HAL_3DRESOURCE_ENTRY *entry = nullptr; for (uint32_t i = 0; i < state->cmDeviceParam.max3DSurfaceTableSize; i++) { if (Mos_ResourceIsNull(&state->surf3DTable[i].osResource)) { entry = &state->surf3DTable[i]; param->handle = (uint32_t)i; break; } } if (!entry) { eStatus = MOS_STATUS_INVALID_PARAMETER; CM_ASSERTMESSAGE("3D surface table is full"); return eStatus; } CM_CHK_NULL_GOTOFINISH_MOSERROR(osInterface); osInterface->pfnResetResource(&entry->osResource); // Resets the Resource MOS_ALLOC_GFXRES_PARAMS alloc_params; MOS_ZeroMemory(&alloc_params, sizeof(alloc_params)); alloc_params.Type = MOS_GFXRES_VOLUME; alloc_params.TileType = MOS_TILE_Y; alloc_params.dwWidth = param->width; alloc_params.dwHeight = param->height; alloc_params.dwDepth = param->depth; alloc_params.pSystemMemory = param->data; alloc_params.Format = param->format; alloc_params.pBufName = "CmSurface3D"; CM_CHK_HRESULT_GOTOFINISH_MOSERROR(osInterface->pfnAllocateResource( osInterface, &alloc_params, &entry->osResource)); entry->width = param->width; entry->height = param->height; entry->depth = param->depth; entry->format = param->format; if (state->advExecutor) { entry->surfStateMgr = state->advExecutor->Create3DStateMgr(&entry->osResource); state->advExecutor->Set2Dor3DOrigDimension(entry->surfStateMgr, entry->width, entry->height, entry->depth); } finish: return eStatus; } //*----------------------------------------------------------------------------- //| Purpose: Frees the resource and removes from the table //| Returns: Result of the operation. //*----------------------------------------------------------------------------- MOS_STATUS HalCm_Free3DResource( PCM_HAL_STATE state, // [in] Pointer to CM State uint32_t handle) // [in] Pointer to Buffer Param { MOS_STATUS eStatus; PCM_HAL_3DRESOURCE_ENTRY entry; PMOS_INTERFACE osInterface; MOS_GFXRES_FREE_FLAGS resFreeFlags = {0}; resFreeFlags.AssumeNotInUse = 1; eStatus = MOS_STATUS_SUCCESS; osInterface = state->osInterface; // Get the Buffer Entry CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_Get3DResourceEntry(state, handle, &entry)); if (state->advExecutor) { state->advExecutor->Delete2Dor3DStateMgr(entry->surfStateMgr); } osInterface->pfnFreeResourceWithFlag(osInterface, &entry->osResource, resFreeFlags.Value); osInterface->pfnResetResourceAllocationIndex(osInterface, &entry->osResource); finish: return eStatus; } //*----------------------------------------------------------------------------- //| Purpose: Lock the resource and return //| Returns: Result of the operation. //*----------------------------------------------------------------------------- MOS_STATUS HalCm_Lock3DResource( PCM_HAL_STATE state, // [in] Pointer to CM State PCM_HAL_3DRESOURCE_PARAM param) // [in] Pointer to 3D Param { MOS_STATUS eStatus = MOS_STATUS_SUCCESS; PCM_HAL_3DRESOURCE_ENTRY entry; MOS_LOCK_PARAMS lockFlags; RENDERHAL_GET_SURFACE_INFO info; PMOS_INTERFACE osInterface = nullptr; MOS_SURFACE surface; // Get the 3D Resource Entry CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_Get3DResourceEntry(state, param->handle, &entry)); if ((param->lockFlag != CM_HAL_LOCKFLAG_READONLY) && (param->lockFlag != CM_HAL_LOCKFLAG_WRITEONLY) ) { CM_ASSERTMESSAGE("Invalid lock flag!"); eStatus = MOS_STATUS_UNKNOWN; goto finish; } // Get resource information MOS_ZeroMemory(&surface, sizeof(surface)); surface.OsResource = entry->osResource; surface.Format = Format_Invalid; osInterface = state->osInterface; MOS_ZeroMemory(&info, sizeof(RENDERHAL_GET_SURFACE_INFO)); CM_CHK_MOSSTATUS_GOTOFINISH(RenderHal_GetSurfaceInfo( osInterface, &info, &surface)); param->pitch = surface.dwPitch; param->qpitch = surface.dwQPitch; param->qpitchEnabled = state->cmHalInterface->IsSurf3DQpitchSupportedbyHw(); // Lock the resource MOS_ZeroMemory(&lockFlags, sizeof(MOS_LOCK_PARAMS)); if (param->lockFlag == CM_HAL_LOCKFLAG_READONLY) { lockFlags.ReadOnly = true; } else { lockFlags.WriteOnly = true; } lockFlags.ForceCached = true; param->data = osInterface->pfnLockResource( osInterface, &entry->osResource, &lockFlags); CM_CHK_NULL_GOTOFINISH_MOSERROR(param->data); finish: return eStatus; } //*----------------------------------------------------------------------------- //| Purpose: Unlock the resource and return //| Returns: Result of the operation. //*----------------------------------------------------------------------------- MOS_STATUS HalCm_Unlock3DResource( PCM_HAL_STATE state, // [in] Pointer to CM State PCM_HAL_3DRESOURCE_PARAM param) // [in] Pointer to 3D Param { MOS_STATUS eStatus; PCM_HAL_3DRESOURCE_ENTRY entry; PMOS_INTERFACE osInterface; eStatus = MOS_STATUS_SUCCESS; osInterface = state->osInterface; // Get the 3D Resource Entry CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_Get3DResourceEntry(state, param->handle, &entry)); // Lock the resource CM_CHK_HRESULT_GOTOFINISH_MOSERROR(osInterface->pfnUnlockResource(osInterface, &entry->osResource)); finish: return eStatus; } MOS_STATUS HalCm_SetCompressionMode( PCM_HAL_STATE state, CM_HAL_SURFACE2D_COMPRESSIOM_PARAM mmcParam) { MOS_STATUS eStatus = MOS_STATUS_SUCCESS; PMOS_INTERFACE osInterface = state->osInterface; PCM_HAL_SURFACE2D_ENTRY entry; // Get the 2D Resource Entry CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_GetSurface2DEntry(state, mmcParam.handle, &entry)); //set compression bit passed down CM_CHK_MOSSTATUS_GOTOFINISH(osInterface->pfnSetMemoryCompressionMode(osInterface, &(entry->osResource), (MOS_MEMCOMP_STATE)mmcParam.mmcMode)); finish: return eStatus; } MOS_STATUS HalCm_SetL3Cache( const L3ConfigRegisterValues *l3Values, PCmHalL3Settings cmHalL3Cache ) { MOS_STATUS eStatus = MOS_STATUS_SUCCESS; // in legacy platforms, we map: // ConfigRegister0->SqcReg1 // ConfigRegister1->CntlReg2 // ConfigRegister2->CntlReg3 // ConfigRegister3->CntlReg CM_CHK_NULL_GOTOFINISH_MOSERROR( cmHalL3Cache ); CM_CHK_NULL_GOTOFINISH_MOSERROR(l3Values); cmHalL3Cache->overrideSettings = (l3Values->config_register0 || l3Values->config_register1 || l3Values->config_register2 || l3Values->config_register3 ); cmHalL3Cache->cntlRegOverride = (l3Values->config_register3 != 0); cmHalL3Cache->cntlReg2Override = (l3Values->config_register1 != 0); cmHalL3Cache->cntlReg3Override = (l3Values->config_register2 != 0); cmHalL3Cache->sqcReg1Override = (l3Values->config_register0 != 0); cmHalL3Cache->cntlReg = l3Values->config_register3; cmHalL3Cache->cntlReg2 = l3Values->config_register1; cmHalL3Cache->cntlReg3 = l3Values->config_register2; cmHalL3Cache->sqcReg1 = l3Values->config_register0; finish: return MOS_STATUS_SUCCESS; } //*----------------------------------------------------------------------------- //| Purpose: Set Cap values //| Returns: Result of the operation. //*----------------------------------------------------------------------------- MOS_STATUS HalCm_SetCaps( PCM_HAL_STATE state, PCM_HAL_MAX_SET_CAPS_PARAM setCapsParam) { MOS_STATUS eStatus = MOS_STATUS_SUCCESS; CM_CHK_NULL_GOTOFINISH_MOSERROR(state); CM_CHK_NULL_GOTOFINISH_MOSERROR(setCapsParam); CM_CHK_NULL_GOTOFINISH_MOSERROR(state->renderHal); CM_CHK_NULL_GOTOFINISH_MOSERROR(state->renderHal->pHwCaps) switch (setCapsParam->type) { case CM_SET_MAX_HW_THREADS: if( setCapsParam->maxValue <= 0 || setCapsParam->maxValue > state->renderHal->pHwCaps->dwMaxThreads ) { eStatus = MOS_STATUS_UNKNOWN; goto finish; } else { state->maxHWThreadValues.apiValue = (setCapsParam->maxValue == 0) ? 0: MOS_MAX(setCapsParam->maxValue, state->cmHalInterface->GetSmallestMaxThreadNum()); } break; case CM_SET_HW_L3_CONFIG: eStatus = state->cmHalInterface->SetL3CacheConfig( &setCapsParam->l3CacheValues, &state->l3Settings ); break; default: eStatus = MOS_STATUS_UNKNOWN; goto finish; } finish: return eStatus; } //*----------------------------------------------------------------------------- //| Purpose: Task sets the power option which will be used by this task //| Returns: Result of the operation. //*----------------------------------------------------------------------------- MOS_STATUS HalCm_SetPowerOption( PCM_HAL_STATE state, PCM_POWER_OPTION powerOption ) { if (state->cmHalInterface->IsOverridePowerOptionPerGpuContext()) { CM_NORMALMESSAGE("WARNING: Deprecated function due to per context SSEU overriding is enabled.\n"); return MOS_STATUS_SUCCESS; } MOS_SecureMemcpy( &state->powerOption, sizeof( state->powerOption ), powerOption, sizeof( state->powerOption ) ); return MOS_STATUS_SUCCESS; } //*----------------------------------------------------------------------------- // Purpose: Get the time in ns from QueryPerformanceCounter // Returns: Result of the operation //*----------------------------------------------------------------------------- MOS_STATUS HalCm_GetGlobalTime(LARGE_INTEGER *globalTime) { if(globalTime == nullptr) { return MOS_STATUS_NULL_POINTER; } if (MosUtilities::MosQueryPerformanceCounter((uint64_t *)&(globalTime->QuadPart)) == false) { return MOS_STATUS_UNKNOWN; } return MOS_STATUS_SUCCESS; } //*----------------------------------------------------------------------------- // Purpose: Convert time from nanosecond to QPC time // Returns: Result of the operation //*----------------------------------------------------------------------------- MOS_STATUS HalCm_ConvertToQPCTime(uint64_t nanoseconds, LARGE_INTEGER *qpcTime) { LARGE_INTEGER perfFreq; if(qpcTime == nullptr) { return MOS_STATUS_NULL_POINTER; } if (MosUtilities::MosQueryPerformanceFrequency((uint64_t*)&perfFreq.QuadPart) == false) { return MOS_STATUS_UNKNOWN; } qpcTime->QuadPart = (uint64_t)(nanoseconds * perfFreq.QuadPart / 1000000000.0); return MOS_STATUS_SUCCESS; } //------------------------------------------------------------------------------ //| Purpose: Halcm updates power state to hw state //| Returns: //------------------------------------------------------------------------------ MOS_STATUS HalCm_UpdatePowerOption( PCM_HAL_STATE state, PCM_POWER_OPTION powerOption ) { MOS_STATUS eStatus = MOS_STATUS_SUCCESS; if (state->cmHalInterface->IsOverridePowerOptionPerGpuContext()) { CM_NORMALMESSAGE("WARNING: Deprecated function due to per context SSEU overriding is enabled.\n"); return MOS_STATUS_SUCCESS; } PRENDERHAL_INTERFACE renderHal = state->renderHal; RENDERHAL_POWEROPTION renderPowerOption; renderPowerOption.nSlice = (uint8_t)powerOption->nSlice; renderPowerOption.nSubSlice = (uint8_t)powerOption->nSubSlice; renderPowerOption.nEU = (uint8_t)powerOption->nEU; // option set in CM create device to use slice shutdown for life of CM device ( override previous value if necessary ) if ( state->requestSingleSlice == true ) { renderPowerOption.nSlice = 1; } renderHal->pfnSetPowerOptionMode( renderHal, &renderPowerOption ); return eStatus; } MOS_STATUS HalCm_InitPerfTagIndexMap(PCM_HAL_STATE cmState) { using namespace std; MOS_STATUS eStatus = MOS_STATUS_SUCCESS; CM_ASSERT(cmState); for (int i = 0; i < MAX_COMBINE_NUM_IN_PERFTAG; i++) { cmState->currentPerfTagIndex[i] = 1; #if MOS_MESSAGES_ENABLED cmState->perfTagIndexMap[i] = MosUtilities::MosNewUtil >(__FUNCTION__, __FILE__, __LINE__); #else cmState->perfTagIndexMap[i] = MosUtilities::MosNewUtil >(); #endif CM_CHK_NULL_GOTOFINISH_MOSERROR(cmState->perfTagIndexMap[i]); } cmState->perfTagIndexMap[0]->insert(pair("surfaceCopy_read_NV12_32x32", GPUCOPY_READ_PERFTAG_INDEX)); cmState->perfTagIndexMap[0]->insert(pair("surfaceCopy_read_NV12_aligned_32x32", GPUCOPY_READ_PERFTAG_INDEX)); cmState->perfTagIndexMap[0]->insert(pair("surfaceCopy_read_32x32", GPUCOPY_READ_PERFTAG_INDEX)); cmState->perfTagIndexMap[0]->insert(pair("surfaceCopy_read_aligned_32x32", GPUCOPY_READ_PERFTAG_INDEX)); cmState->perfTagIndexMap[0]->insert(pair("surfaceCopy_write_NV12_32x32", GPUCOPY_WRITE_PERFTAG_INDEX)); cmState->perfTagIndexMap[0]->insert(pair("surfaceCopy_write_32x32", GPUCOPY_WRITE_PERFTAG_INDEX)); cmState->perfTagIndexMap[0]->insert(pair("SurfaceCopy_2DTo2D_NV12_32x32", GPUCOPY_G2G_PERFTAG_INDEX)); cmState->perfTagIndexMap[0]->insert(pair("SurfaceCopy_2DTo2D_32x32", GPUCOPY_G2G_PERFTAG_INDEX)); cmState->perfTagIndexMap[0]->insert(pair("SurfaceCopy_BufferToBuffer_4k", GPUCOPY_C2C_PERFTAG_INDEX)); cmState->perfTagIndexMap[0]->insert(pair("SurfaceCopy_BufferToBuffer_4k", GPUCOPY_C2C_PERFTAG_INDEX)); cmState->perfTagIndexMap[0]->insert(pair("surfaceCopy_set_NV12", GPUINIT_PERFTAG_INDEX)); cmState->perfTagIndexMap[0]->insert(pair("surfaceCopy_set", GPUINIT_PERFTAG_INDEX)); finish: return eStatus; } MOS_STATUS HalCm_DeleteFromStateBufferList( PCM_HAL_STATE state, void *kernelPtr ) { MOS_STATUS result = MOS_STATUS_SUCCESS; state->state_buffer_list_ptr->erase( kernelPtr ); return result; } PRENDERHAL_MEDIA_STATE HalCm_GetMediaStatePtrForKernel( PCM_HAL_STATE state, void *kernelPtr ) { if ( state->state_buffer_list_ptr->find( kernelPtr ) != state->state_buffer_list_ptr->end() ) { return ( *state->state_buffer_list_ptr )[ kernelPtr ].mediaStatePtr; } else { return nullptr; } } uint64_t HalCm_GetStateBufferVAPtrForSurfaceIndex( PCM_HAL_STATE state, uint32_t surfIndex ) { for ( auto listItem = state->state_buffer_list_ptr->begin(); listItem != state->state_buffer_list_ptr->end(); listItem++ ) { if ( listItem->second.stateBufferIndex == surfIndex ) { return listItem->second.stateBufferVaPtr; } } return 0; } PRENDERHAL_MEDIA_STATE HalCm_GetMediaStatePtrForSurfaceIndex( PCM_HAL_STATE state, uint32_t surfIndex ) { for ( auto listItem = state->state_buffer_list_ptr->begin(); listItem != state->state_buffer_list_ptr->end(); listItem++ ) { if ( listItem->second.stateBufferIndex == surfIndex ) { return listItem->second.mediaStatePtr; } } return nullptr; } uint64_t HalCm_GetStateBufferVAPtrForMediaStatePtr( PCM_HAL_STATE state, PRENDERHAL_MEDIA_STATE mediaStatePtr ) { for ( auto listItem = state->state_buffer_list_ptr->begin(); listItem != state->state_buffer_list_ptr->end(); listItem++ ) { if ( listItem->second.mediaStatePtr == mediaStatePtr ) { return listItem->second.stateBufferVaPtr; } } return 0; } uint32_t HalCm_GetStateBufferSizeForKernel( PCM_HAL_STATE state, void *kernelPtr ) { if ( state->state_buffer_list_ptr->find( kernelPtr ) != state->state_buffer_list_ptr->end() ) { return ( *state->state_buffer_list_ptr )[ kernelPtr ].stateBufferSize; } else { return 0; } } CM_STATE_BUFFER_TYPE HalCm_GetStateBufferTypeForKernel( PCM_HAL_STATE state, void *kernelPtr ) { if ( state->state_buffer_list_ptr->find( kernelPtr ) != state->state_buffer_list_ptr->end() ) { return ( *state->state_buffer_list_ptr )[ kernelPtr ].stateBufferType; } else { return CM_STATE_BUFFER_NONE; } } static void LoadUserFeatures(CM_HAL_STATE *halState, MOS_GPUCTX_CREATOPTIONS *createOptions) { #if (_DEBUG || _RELEASE_INTERNAL) MOS_USER_FEATURE_VALUE_DATA user_feature_data; MOS_ZeroMemory(&user_feature_data, sizeof(user_feature_data)); MOS_STATUS result = MOS_UserFeature_ReadValue_ID( nullptr, __MEDIA_USER_FEATURE_VALUE_MDF_FORCE_RAMODE, &user_feature_data, halState->osInterface->pOsContext); if (MOS_STATUS_SUCCESS == result && user_feature_data.i32Data == 1) { createOptions->RAMode = 1; } MOS_USER_FEATURE_VALUE_WRITE_DATA userFeatureWriteData; userFeatureWriteData = __NULL_USER_FEATURE_VALUE_WRITE_DATA__; userFeatureWriteData.Value.i32Data = createOptions->RAMode; userFeatureWriteData.ValueID = __MEDIA_USER_FEATURE_VALUE_MDF_FORCE_RAMODE; MOS_UserFeature_WriteValues_ID(nullptr, &userFeatureWriteData, 1, halState->osInterface->pOsContext); #endif return; } MOS_STATUS HalCm_CreateGPUContext( PCM_HAL_STATE state, MOS_GPU_CONTEXT gpuContext, MOS_GPU_NODE gpuNode, PMOS_GPUCTX_CREATOPTIONS pMosGpuContextCreateOption) { MOS_STATUS eStatus = MOS_STATUS_SUCCESS; LoadUserFeatures(state, pMosGpuContextCreateOption); // Create Compute Context on Compute Node CM_CHK_HRESULT_GOTOFINISH_MOSERROR(state->osInterface->pfnCreateGpuContext( state->osInterface, gpuContext, gpuNode, pMosGpuContextCreateOption)); // Register Compute Context with the Batch Buffer completion event CM_CHK_HRESULT_GOTOFINISH_MOSERROR(state->osInterface->pfnRegisterBBCompleteNotifyEvent( state->osInterface, gpuContext)); finish: return eStatus; } GPU_CONTEXT_HANDLE HalCm_CreateGpuComputeContext(CM_HAL_STATE *state, MOS_GPUCTX_CREATOPTIONS *createOptions) { LoadUserFeatures(state, createOptions); GPU_CONTEXT_HANDLE context_handle = state->osInterface->pfnCreateGpuComputeContext( state->osInterface, MOS_GPU_CONTEXT_CM_COMPUTE, createOptions); if (MOS_GPU_CONTEXT_INVALID_HANDLE != context_handle) { state->osInterface->pfnRegisterBBCompleteNotifyEvent( state->osInterface, MOS_GPU_CONTEXT_CM_COMPUTE); } return context_handle; } uint32_t HalCm_SetGpuContext(CM_HAL_STATE *halState, MOS_GPU_CONTEXT contextName, uint32_t streamIndex, GPU_CONTEXT_HANDLE contextHandle) { uint32_t old_stream_idx = halState->osInterface->streamIndex; halState->osInterface->streamIndex = streamIndex; MOS_STATUS result = MOS_STATUS_SUCCESS; if (MOS_GPU_CONTEXT_INVALID_HANDLE == contextHandle) { result = halState->osInterface->pfnSetGpuContext(halState->osInterface, contextName); } else { result = halState->osInterface->pfnSetGpuContextFromHandle( halState->osInterface, contextName, contextHandle); } if (MOS_STATUS_SUCCESS != result) { halState->osInterface->streamIndex = old_stream_idx; return INVALID_STREAM_INDEX; } return old_stream_idx; } MOS_STATUS HalCm_SelectSyncBuffer(CM_HAL_STATE *halState, uint32_t bufferIdx) { if (bufferIdx >= halState->cmDeviceParam.maxBufferTableSize) { halState->syncBuffer = nullptr; return MOS_STATUS_SUCCESS; } CM_HAL_BUFFER_ENTRY *entry = halState->bufferTable + bufferIdx; halState->syncBuffer = &entry->osResource; MOS_INTERFACE *os_interface = halState->osInterface; return os_interface->pfnRegisterResource(os_interface, halState->syncBuffer, true, true); } //*----------------------------------------------------------------------------- //| Purpose: Creates instance of HAL CM State //| Returns: Result of the operation //| Note: Caller must call pfnAllocate to allocate all HalCm/Mhw states and objects. //| Caller MUST call HalCm_Destroy to destroy the instance //*----------------------------------------------------------------------------- MOS_STATUS HalCm_Create( PMOS_CONTEXT osDriverContext, // [in] OS Driver Context PCM_HAL_CREATE_PARAM param, // [in] Create Param PCM_HAL_STATE *cmState) // [out] double pointer to CM State { MOS_STATUS eStatus; PCM_HAL_STATE state = nullptr; uint32_t numCmdBuffers = 0; MhwInterfaces *mhwInterfaces = nullptr; MhwInterfaces::CreateParams params; MOS_GPUCTX_CREATOPTIONS createOption; //----------------------------------------- CM_ASSERT(osDriverContext); CM_ASSERT(param); CM_ASSERT(cmState); //----------------------------------------- eStatus = MOS_STATUS_SUCCESS; // Allocate State structure state = (PCM_HAL_STATE)MOS_AllocAndZeroMemory(sizeof(CM_HAL_STATE)); CM_CHK_NULL_GOTOFINISH_MOSERROR(state); // Allocate/Initialize OS Interface state->osInterface = (PMOS_INTERFACE) MOS_AllocAndZeroMemory(sizeof(MOS_INTERFACE)); CM_CHK_NULL_GOTOFINISH_MOSERROR(state->osInterface); state->osInterface->bDeallocateOnExit = true; CM_CHK_HRESULT_GOTOFINISH_MOSERROR(Mos_InitInterface(state->osInterface, osDriverContext, COMPONENT_CM)); #if (_RELEASE_INTERNAL || _DEBUG) #if defined(CM_DIRECT_GUC_SUPPORT) state->osInterface->m_pWorkQueueMngr = new CMRTWorkQueueMngr(); #endif #endif state->osInterface->pfnGetPlatform(state->osInterface, &state->platform); state->skuTable = state->osInterface->pfnGetSkuTable(state->osInterface); state->waTable = state->osInterface->pfnGetWaTable (state->osInterface); // Create VEBOX Context createOption.CmdBufferNumScale = MOS_GPU_CONTEXT_CREATE_DEFAULT; CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_CreateGPUContext( state, MOS_GPU_CONTEXT_VEBOX, MOS_GPU_NODE_VE, &createOption)); state->osInterface->pfnSetGpuContext(state->osInterface, MOS_GPU_CONTEXT_VEBOX); // Allocate/Initialize CM Rendering Interface state->renderHal = (PRENDERHAL_INTERFACE_LEGACY) MOS_AllocAndZeroMemory(sizeof(RENDERHAL_INTERFACE_LEGACY)); CM_CHK_NULL_GOTOFINISH_MOSERROR(state->renderHal); state->dshEnabled = param->dynamicStateHeap; state->renderHal->bDynamicStateHeap = state->dshEnabled; if (state->dshEnabled) { CM_CHK_MOSSTATUS_GOTOFINISH(RenderHal_InitInterface_Dynamic(state->renderHal, &state->cpInterface, state->osInterface)); } else { CM_CHK_MOSSTATUS_GOTOFINISH(RenderHal_InitInterface_Legacy(state->renderHal, &state->cpInterface, state->osInterface)); } // Allocate/Initialize VEBOX Interface if (!param->disableVebox) { CmSafeMemSet(¶ms, 0, sizeof(params)); params.Flags.m_vebox = 1; mhwInterfaces = MhwInterfaces::CreateFactory(params, state->osInterface); if (mhwInterfaces) { state->veboxInterface = mhwInterfaces->m_veboxInterface; // MhwInterfaces always create CP and MI interfaces, so we have to delete those we don't need. MOS_Delete(mhwInterfaces->m_miInterface); state->osInterface->pfnDeleteMhwCpInterface(mhwInterfaces->m_cpInterface); mhwInterfaces->m_cpInterface = nullptr; MOS_Delete(mhwInterfaces); CM_CHK_NULL_GOTOFINISH_MOSERROR(state->veboxInterface); } else { CM_ASSERTMESSAGE("Allocate MhwInterfaces failed"); HalCm_Destroy(state); *cmState = nullptr; return MOS_STATUS_NO_SPACE; } } else { state->veboxInterface = nullptr; } // set IsMDFLoad to distinguish MDF context from other Media Contexts state->renderHal->IsMDFLoad = true; // disable YV12SinglePass as CMRT & compiler don't support it state->renderHal->bEnableYV12SinglePass = false; state->cmDeviceParam.maxKernelBinarySize = CM_KERNEL_BINARY_BLOCK_SIZE; // set if the new sampler heap management is used or not // currently new sampler heap management depends on DSH if (state->dshEnabled) { state->useNewSamplerHeap = true; } else { state->useNewSamplerHeap = false; } //Get Max Scratch Space Size if( param->disableScratchSpace) { state->cmDeviceParam.maxPerThreadScratchSpaceSize = 0; } else { //Gen7_5 + : (MaxScratchSpaceSize + 1) *16k if(param->scratchSpaceSize == CM_DEVICE_CONFIG_SCRATCH_SPACE_SIZE_DEFAULT) { //By default, 128K for HSW state->cmDeviceParam.maxPerThreadScratchSpaceSize = 8 * CM_DEVICE_CONFIG_SCRATCH_SPACE_SIZE_16K_STEP; } else { state->cmDeviceParam.maxPerThreadScratchSpaceSize = (param->scratchSpaceSize)* CM_DEVICE_CONFIG_SCRATCH_SPACE_SIZE_16K_STEP; } } // Initialize kernel parameters state->kernelParamsRenderHal.pMhwKernelParam = &state->kernelParamsMhw; // Enable SLM in L3 Cache state->l3Settings.enableSlm = true; // Slice shutdown state->requestSingleSlice = param->requestSliceShutdown; //mid thread preemption on/off and SIP debug control state->midThreadPreemptionDisabled = param->disabledMidThreadPreemption; state->kernelDebugEnabled = param->enabledKernelDebug; // init mapping for the state buffer #if MOS_MESSAGES_ENABLED state->state_buffer_list_ptr = MosUtilities::MosNewUtil >(__FUNCTION__, __FILE__, __LINE__); #else state->state_buffer_list_ptr = MosUtilities::MosNewUtil >(); #endif CM_CHK_NULL_GOTOFINISH_MOSERROR( state->state_buffer_list_ptr ); MOS_ZeroMemory(&state->hintIndexes.kernelIndexes, sizeof(uint32_t) * CM_MAX_TASKS_EU_SATURATION); MOS_ZeroMemory(&state->hintIndexes.dispatchIndexes, sizeof(uint32_t) * CM_MAX_TASKS_EU_SATURATION); // get the global media profiler state->perfProfiler = MediaPerfProfiler::Instance(); CM_CHK_NULL_GOTOFINISH_MOSERROR(state->perfProfiler); CM_CHK_MOSSTATUS_GOTOFINISH(state->perfProfiler->Initialize((void*)state, state->osInterface)); state->criticalSectionDSH = MOS_New(CMRT_UMD::CSync); CM_CHK_NULL_GOTOFINISH_MOSERROR(state->criticalSectionDSH); state->cmDeviceParam.maxKernelsPerTask = CM_MAX_KERNELS_PER_TASK; state->cmDeviceParam.maxSamplerTableSize = CM_MAX_SAMPLER_TABLE_SIZE; state->cmDeviceParam.maxSampler8x8TableSize = state->renderHal->pHwSizes->dwSizeSampler8x8Table; state->cmDeviceParam.maxBufferTableSize = CM_MAX_BUFFER_SURFACE_TABLE_SIZE; state->cmDeviceParam.max2DSurfaceUPTableSize = CM_MAX_2D_SURFACE_UP_TABLE_SIZE; state->cmDeviceParam.max2DSurfaceTableSize = CM_MAX_2D_SURFACE_TABLE_SIZE; state->cmDeviceParam.max3DSurfaceTableSize = CM_MAX_3D_SURFACE_TABLE_SIZE; state->cmDeviceParam.maxTasks = param->maxTaskNumber; state->cmDeviceParam.maxAvsSamplers = CM_MAX_AVS_SAMPLER_SIZE; state->cmDeviceParam.maxGshKernelEntries = param->kernelBinarySizeinGSH / (CM_32K); if (state->dshEnabled) { // Initialize Kernel Cache Hit/Miss counters state->dshKernelCacheMiss = 0; state->dshKernelCacheHit = 0; } // Setup Function pointers state->pfnCmAllocate = HalCm_Allocate; state->pfnGetMaxValues = HalCm_GetMaxValues; state->pfnGetMaxValuesEx = HalCm_GetMaxValuesEx; state->pfnExecuteTask = HalCm_ExecuteTask; state->pfnExecuteGroupTask = HalCm_ExecuteGroupTask; state->pfnExecuteHintsTask = HalCm_ExecuteHintsTask; state->pfnRegisterSampler = HalCm_RegisterSampler; state->pfnUnRegisterSampler = HalCm_UnRegisterSampler; state->pfnRegisterSampler8x8 = HalCm_RegisterSampler8x8; state->pfnUnRegisterSampler8x8 = HalCm_UnRegisterSampler8x8; state->pfnFreeBuffer = HalCm_FreeBuffer; state->pfnLockBuffer = HalCm_LockBuffer; state->pfnUnlockBuffer = HalCm_UnlockBuffer; state->pfnFreeSurface2DUP = HalCm_FreeSurface2DUP; state->pfnGetSurface2DTileYPitch = HalCm_GetSurface2DTileYPitch; state->pfnSet2DSurfaceStateParam = HalCm_Set2DSurfaceStateParam; state->pfnSetBufferSurfaceStatePara = HalCm_SetBufferSurfaceStateParameters; state->pfnSetSurfaceMOCS = HalCm_SetSurfaceMOCS; /************************************************************/ state->pfnAllocateSurface2D = HalCm_AllocateSurface2D; state->pfnAllocate3DResource = HalCm_AllocateSurface3D; state->pfnFreeSurface2D = HalCm_FreeSurface2D; state->pfnLock2DResource = HalCm_Lock2DResource; state->pfnUnlock2DResource = HalCm_Unlock2DResource; state->pfnSetCompressionMode = HalCm_SetCompressionMode; /************************************************************/ state->pfnFree3DResource = HalCm_Free3DResource; state->pfnLock3DResource = HalCm_Lock3DResource; state->pfnUnlock3DResource = HalCm_Unlock3DResource; state->pfnSetCaps = HalCm_SetCaps; state->pfnSetPowerOption = HalCm_SetPowerOption; state->pfnUpdatePowerOption = HalCm_UpdatePowerOption; state->pfnSendMediaWalkerState = HalCm_SendMediaWalkerState; state->pfnSendGpGpuWalkerState = HalCm_SendGpGpuWalkerState; state->pfnSetSurfaceReadFlag = HalCm_SetSurfaceReadFlag; state->pfnSetVtuneProfilingFlag = HalCm_SetVtuneProfilingFlag; state->pfnExecuteVeboxTask = HalCm_ExecuteVeboxTask; state->pfnGetTaskSyncLocation = HalCm_GetTaskSyncLocation; state->pfnGetGlobalTime = HalCm_GetGlobalTime; state->pfnConvertToQPCTime = HalCm_ConvertToQPCTime; state->pfnSyncOnResource = HalCm_SyncOnResource; state->pfnDeleteFromStateBufferList = HalCm_DeleteFromStateBufferList; state->pfnGetMediaStatePtrForKernel = HalCm_GetMediaStatePtrForKernel; state->pfnGetStateBufferVAPtrForSurfaceIndex = HalCm_GetStateBufferVAPtrForSurfaceIndex; state->pfnGetMediaStatePtrForSurfaceIndex = HalCm_GetMediaStatePtrForSurfaceIndex; state->pfnGetStateBufferVAPtrForMediaStatePtr = HalCm_GetStateBufferVAPtrForMediaStatePtr; state->pfnGetStateBufferSizeForKernel = HalCm_GetStateBufferSizeForKernel; state->pfnGetStateBufferTypeForKernel = HalCm_GetStateBufferTypeForKernel; state->pfnCreateGPUContext = HalCm_CreateGPUContext; state->pfnCreateGpuComputeContext = HalCm_CreateGpuComputeContext; state->pfnSetGpuContext = HalCm_SetGpuContext; state->pfnSelectSyncBuffer = HalCm_SelectSyncBuffer; state->pfnDSHUnregisterKernel = HalCm_DSH_UnregisterKernel; state->pfnUpdateBuffer = HalCm_UpdateBuffer; state->pfnUpdateSurface2D = HalCm_UpdateSurface2D; //============== HalCm_OsInitInterface(state); state->osInterface->pfnInitCmInterface(state); HalCm_InitPerfTagIndexMap(state); state->maxHWThreadValues.userFeatureValue = 0; state->maxHWThreadValues.apiValue = 0; HalCm_GetUserFeatureSettings(state); #if MDF_COMMAND_BUFFER_DUMP HalCm_InitDumpCommandBuffer(state); state->pfnInitDumpCommandBuffer = HalCm_InitDumpCommandBuffer; state->pfnDumpCommadBuffer = HalCm_DumpCommadBuffer; #endif //MDF_COMMAND_BUFFER_DUMP #if MDF_CURBE_DATA_DUMP HalCm_InitDumpCurbeData(state); #endif #if MDF_SURFACE_CONTENT_DUMP HalCm_InitSurfaceDump(state); #endif #if MDF_SURFACE_STATE_DUMP HalCm_InitDumpSurfaceState(state); state->pfnInitDumpSurfaceState = HalCm_InitDumpSurfaceState; state->pfnDumpSurfaceState = HalCm_DumpSurfaceState; #endif #if MDF_INTERFACE_DESCRIPTOR_DATA_DUMP HalCm_InitDumpInterfaceDescriporData(state); #endif state->cmHalInterface = CMHalDevice::CreateFactory(state); CM_CHK_NULL_GOTOFINISH_MOSERROR(state->cmHalInterface); if (param->refactor) { state->refactor = true; } else { state->refactor = false; } state->requestCustomGpuContext = param->requestCustomGpuContext; #if (_DEBUG || _RELEASE_INTERNAL) { MOS_USER_FEATURE_VALUE_DATA userFeatureData; MOS_ZeroMemory(&userFeatureData, sizeof(userFeatureData)); MOS_UserFeature_ReadValue_ID( nullptr, __MEDIA_USER_FEATURE_VALUE_MDF_FORCE_EXECUTION_PATH_ID, &userFeatureData, state->osInterface->pOsContext); if (userFeatureData.i32Data == 1) { state->refactor = false; } else if (userFeatureData.i32Data == 2) { state->refactor = true; state->cmHalInterface->SetFastPathByDefault(true); } FILE *fp1 = nullptr; MosUtilities::MosSecureFileOpen(&fp1, "refactor.key", "r"); if (fp1 != nullptr) { state->refactor = true; state->cmHalInterface->SetFastPathByDefault(true); fclose(fp1); } FILE *fp2 = nullptr; MosUtilities::MosSecureFileOpen(&fp2, "origin.key", "r"); if (fp2 != nullptr) { state->refactor = false; fclose(fp2); } } #endif if (state->refactor) { CM_NORMALMESSAGE("Info: Fast path is enabled!\n"); } else { CM_NORMALMESSAGE("Info: Fast path is disabled!\n"); } finish: if (eStatus != MOS_STATUS_SUCCESS) { HalCm_Destroy(state); *cmState = nullptr; } else { *cmState = state; } return eStatus; } //*----------------------------------------------------------------------------- //| Purpose: Destroys instance of HAL CM State //| Returns: N/A //*----------------------------------------------------------------------------- void HalCm_Destroy( PCM_HAL_STATE state) // [in] Pointer to CM State { MOS_STATUS eStatus; int32_t i; if (state) { //Delete CmHal Interface MosSafeDelete(state->cmHalInterface); if (state->osInterface) { state->osInterface->pfnDeleteMhwCpInterface(state->cpInterface); state->cpInterface = nullptr; } else { CM_ASSERTMESSAGE("Failed to destroy cpInterface."); } MosSafeDelete(state->state_buffer_list_ptr); MosSafeDelete(state->criticalSectionDSH); // Delete the unified media profiler if (state->perfProfiler) { MediaPerfProfiler::Destroy(state->perfProfiler, (void*)state, state->osInterface); state->perfProfiler = nullptr; } // Delete Batch Buffers if (state->batchBuffers) { for (i=0; i < state->numBatchBuffers; i++) { if (!Mos_ResourceIsNull(&state->batchBuffers[i].OsResource)) { eStatus = (MOS_STATUS)state->renderHal->pfnFreeBB( state->renderHal, &state->batchBuffers[i]); CM_ASSERT(eStatus == MOS_STATUS_SUCCESS); } MOS_FreeMemory(state->batchBuffers[i].pPrivateData); } MOS_FreeMemory(state->batchBuffers); state->batchBuffers = nullptr; } // Delete TimeStamp Buffer HalCm_FreeTsResource(state); if ((state->midThreadPreemptionDisabled == false) || (state->kernelDebugEnabled == true)) { // Delete CSR surface HalCm_FreeCsrResource(state); // Delete sip surface HalCm_FreeSipResource(state); } // Delete tracker resource HalCm_FreeTrackerResources(state); // Delete advance executor MOS_Delete(state->advExecutor); // Delete heap manager if (state->renderHal) { MOS_Delete(state->renderHal->dgsheapManager); } if (state->hLibModule) { MosUtilities::MosFreeLibrary(state->hLibModule); state->hLibModule = nullptr; } // Delete RenderHal Interface if (state->renderHal) { if (state->renderHal->pfnDestroy) { state->renderHal->pfnDestroy(state->renderHal); } MOS_FreeMemory(state->renderHal); state->renderHal = nullptr; } // Delete VEBOX Interface if (state->veboxInterface && state->veboxInterface->m_veboxHeap) { state->veboxInterface->DestroyHeap( ); MOS_Delete(state->veboxInterface); state->veboxInterface = nullptr; } // Delete OS Interface if (state->osInterface) { if (state->osInterface->pfnDestroy) { state->osInterface->pfnDestroy(state->osInterface, true); } if (state->osInterface->bDeallocateOnExit) { MOS_FreeMemory(state->osInterface); state->osInterface = nullptr; } } // Delete the TaskParam MOS_FreeMemory(state->taskParam); // Delete the TaskTimeStamp MOS_FreeMemory(state->taskTimeStamp); // Delete Tables MOS_FreeMemory(state->tableMemories); // Delete the pTotalKernelSize table for GSH MOS_FreeMemory(state->totalKernelSize); // Delete the perfTag Map for (int i = 0; i < MAX_COMBINE_NUM_IN_PERFTAG; i++) { MosSafeDelete(state->perfTagIndexMap[i]); } // Delete the state MOS_FreeMemory(state); } } void HalCm_GetUserFeatureSettings( PCM_HAL_STATE cmState ) { #if (_DEBUG || _RELEASE_INTERNAL) PMOS_INTERFACE osInterface = cmState->osInterface; MOS_USER_FEATURE_VALUE_DATA userFeatureData; MOS_ZeroMemory(&userFeatureData, sizeof(userFeatureData)); MOS_UserFeature_ReadValue_ID( nullptr, __MEDIA_USER_FEATURE_VALUE_MDF_MAX_THREAD_NUM_ID, &userFeatureData, cmState->osInterface->pOsContext); if (userFeatureData.i32Data != 0) { uint32_t data = userFeatureData.i32Data; if ((data > 0) && (data <= cmState->renderHal->pHwCaps->dwMaxThreads)) { cmState->maxHWThreadValues.userFeatureValue = data; } } #else UNUSED(cmState); #endif // _DEBUG || _RELEASE_INTERNAL } //*----------------------------------------------------------------------------- //| Purpose: Gathers information about the surface - used by GT-Pin //| Returns: MOS_STATUS_SUCCESS if surface type recognized, S_FAIL otherwise //*----------------------------------------------------------------------------- MOS_STATUS HalCm_GetSurfaceDetails( PCM_HAL_STATE cmState, PCM_HAL_INDEX_PARAM indexParam, uint32_t btIndex, MOS_SURFACE& surface, int16_t globalSurface, PRENDERHAL_SURFACE_STATE_ENTRY surfaceEntry, uint32_t tempPlaneIndex, RENDERHAL_SURFACE_STATE_PARAMS surfaceParam, CM_HAL_KERNEL_ARG_KIND argKind ) { MOS_STATUS eStatus = MOS_STATUS_UNKNOWN; PCM_SURFACE_DETAILS surfaceInfos = nullptr; PCM_SURFACE_DETAILS pgSurfaceInfos = nullptr; PCM_HAL_TASK_PARAM taskParam = cmState->taskParam; uint32_t curKernelIndex = taskParam->curKernelIndex; PMOS_PLANE_OFFSET planeOffset = 0; uint32_t maxEntryNum = 0; MOS_OS_FORMAT tempOsFormat ; CM_SURFACE_BTI_INFO surfBTIInfo; CM_CHK_NULL_GOTOFINISH_MOSERROR(cmState); CM_CHK_NULL_GOTOFINISH_MOSERROR(cmState->cmHalInterface); CM_CHK_NULL_GOTOFINISH_MOSERROR(cmState->osInterface); cmState->cmHalInterface->GetHwSurfaceBTIInfo(&surfBTIInfo); UNUSED(indexParam); if(curKernelIndex+1>taskParam->surfEntryInfoArrays.kernelNum) { eStatus = MOS_STATUS_INVALID_PARAMETER; CM_ASSERTMESSAGE( "Mismatched kernel index: curKernelIndex '%d' vs krnNum '%d'", curKernelIndex,taskParam->surfEntryInfoArrays.kernelNum); goto finish; } surfaceInfos = taskParam->surfEntryInfoArrays.surfEntryInfosArray[curKernelIndex].surfEntryInfos; pgSurfaceInfos = taskParam->surfEntryInfoArrays.surfEntryInfosArray[curKernelIndex].globalSurfInfos; tempOsFormat = (MOS_OS_FORMAT)cmState->osInterface->pfnMosFmtToOsFmt(surface.Format); switch (argKind) { case CM_ARGUMENT_SURFACEBUFFER: if((btIndex >= surfBTIInfo.reservedSurfaceStart) && (btIndex < surfBTIInfo.reservedSurfaceStart + CM_MAX_GLOBAL_SURFACE_NUMBER)) { btIndex = btIndex - surfBTIInfo.reservedSurfaceStart; maxEntryNum = taskParam->surfEntryInfoArrays.surfEntryInfosArray->globalSurfNum; if ( btIndex >= maxEntryNum ) { eStatus = MOS_STATUS_INVALID_PARAMETER; CM_ASSERTMESSAGE( "Array for surface details is full: Max number of entries '%d' and trying to add index '%d'", maxEntryNum, btIndex); goto finish; } MOS_ZeroMemory(&pgSurfaceInfos[btIndex], sizeof(CM_SURFACE_DETAILS)); pgSurfaceInfos[btIndex].width = surface.dwWidth; pgSurfaceInfos[btIndex].format = DDI_FORMAT_UNKNOWN; } else { btIndex = btIndex - surfBTIInfo.reservedSurfaceStart - CM_MAX_GLOBAL_SURFACE_NUMBER; maxEntryNum = taskParam->surfEntryInfoArrays.surfEntryInfosArray->maxEntryNum; if ( btIndex >= maxEntryNum ) { eStatus = MOS_STATUS_INVALID_PARAMETER; CM_ASSERTMESSAGE( "Array for surface details is full: Max number of entries '%d' and trying to add index '%d'", maxEntryNum, btIndex); goto finish; } MOS_ZeroMemory(&surfaceInfos[btIndex], sizeof(CM_SURFACE_DETAILS)); surfaceInfos[btIndex].width = surface.dwWidth; surfaceInfos[btIndex].format = DDI_FORMAT_UNKNOWN; } if (globalSurface < 0) { ++taskParam->surfEntryInfoArrays.surfEntryInfosArray[curKernelIndex].usedIndex; } eStatus = MOS_STATUS_SUCCESS; break; case CM_ARGUMENT_SURFACE2D_UP: case CM_ARGUMENT_SURFACE2D: // VME surface and sampler8x8 called with CM_ARGUMENT_SURFACE2D btIndex = btIndex - surfBTIInfo.reservedSurfaceStart - CM_MAX_GLOBAL_SURFACE_NUMBER; maxEntryNum = taskParam->surfEntryInfoArrays.surfEntryInfosArray->maxEntryNum; if ( btIndex >= maxEntryNum ) { eStatus = MOS_STATUS_INVALID_PARAMETER; CM_ASSERTMESSAGE( "Array for surface details is full: Max number of entries '%d' and trying to add index '%d'", maxEntryNum, btIndex); goto finish; } surfaceInfos[btIndex].width = surfaceEntry->dwWidth; surfaceInfos[btIndex].height = surfaceEntry->dwHeight; surfaceInfos[btIndex].depth = 0; surfaceInfos[btIndex].format = (DdiSurfaceFormat)tempOsFormat; surfaceInfos[btIndex].planeIndex = tempPlaneIndex; surfaceInfos[btIndex].pitch = surfaceEntry->dwPitch; surfaceInfos[btIndex].slicePitch = 0; surfaceInfos[btIndex].surfaceBaseAddress = 0; surfaceInfos[btIndex].tileWalk = surfaceEntry->bTileWalk; surfaceInfos[btIndex].tiledSurface = surfaceEntry->bTiledSurface; if (surfaceEntry->YUVPlane == MHW_U_PLANE || surfaceEntry->YUVPlane == MHW_V_PLANE) { planeOffset = (surfaceEntry->YUVPlane == MHW_U_PLANE) ? &surface.UPlaneOffset : &surface.VPlaneOffset; surfaceInfos[btIndex].yOffset = planeOffset->iYOffset >> 1; if ( argKind == CM_ARGUMENT_SURFACE2D_UP ) { surfaceInfos[btIndex].xOffset = (planeOffset->iXOffset/(uint32_t)sizeof(uint32_t)) >> 2; } else { uint32_t pixelsPerSampleUV = 0; //Get Pixels Per Sample if we use dataport read if(surfaceParam.bWidthInDword_UV) { RenderHal_GetPixelsPerSample(surface.Format, &pixelsPerSampleUV); } else { // If the kernel uses sampler - do not change width (it affects coordinates) pixelsPerSampleUV = 1; } if(pixelsPerSampleUV == 1) { surfaceInfos[btIndex].xOffset = planeOffset->iXOffset >> 2; } else { surfaceInfos[btIndex].xOffset = (planeOffset->iXOffset/(uint32_t)sizeof(uint32_t)) >> 2; } } } else { surfaceInfos[btIndex].xOffset = (surface.YPlaneOffset.iXOffset/(uint32_t)sizeof(uint32_t)) >> 2; surfaceInfos[btIndex].yOffset = surface.YPlaneOffset.iYOffset >> 1; } ++taskParam->surfEntryInfoArrays.surfEntryInfosArray[curKernelIndex].usedIndex; ++tempPlaneIndex; eStatus = MOS_STATUS_SUCCESS; break; case CM_ARGUMENT_SURFACE3D: btIndex = btIndex - surfBTIInfo.normalSurfaceStart - CM_MAX_GLOBAL_SURFACE_NUMBER; maxEntryNum = taskParam->surfEntryInfoArrays.surfEntryInfosArray->maxEntryNum; if ( btIndex >= maxEntryNum ) { eStatus = MOS_STATUS_INVALID_PARAMETER; CM_ASSERTMESSAGE( "Array for surface details is full: Max number of entries '%d' and trying to add index '%d'", maxEntryNum, btIndex); goto finish; } surfaceInfos[btIndex].width = surfaceEntry->dwWidth; surfaceInfos[btIndex].height = surfaceEntry->dwHeight; surfaceInfos[btIndex].depth = surface.dwDepth; surfaceInfos[btIndex].format = (DdiSurfaceFormat)tempOsFormat; surfaceInfos[btIndex].pitch = surfaceEntry->dwPitch; surfaceInfos[btIndex].planeIndex = tempPlaneIndex; surfaceInfos[btIndex].slicePitch = surface.dwSlicePitch; surfaceInfos[btIndex].surfaceBaseAddress = 0; surfaceInfos[btIndex].tileWalk = surfaceEntry->bTileWalk; surfaceInfos[btIndex].tiledSurface = surfaceEntry->bTiledSurface; if (surfaceEntry->YUVPlane == MHW_U_PLANE || surfaceEntry->YUVPlane == MHW_V_PLANE) { planeOffset = (surfaceEntry->YUVPlane == MHW_U_PLANE) ? &surface.UPlaneOffset : &surface.VPlaneOffset; surfaceInfos[btIndex].yOffset = planeOffset->iYOffset >> 1; surfaceInfos[btIndex].xOffset = (planeOffset->iXOffset/(uint32_t)sizeof(uint32_t)) >> 2; } else { surfaceInfos[btIndex].xOffset = (surface.YPlaneOffset.iXOffset/(uint32_t)sizeof(uint32_t)) >> 2; surfaceInfos[btIndex].yOffset = surface.YPlaneOffset.iYOffset >> 1; } ++tempPlaneIndex; ++taskParam->surfEntryInfoArrays.surfEntryInfosArray[curKernelIndex].usedIndex; eStatus = MOS_STATUS_SUCCESS; break; default: break; } finish: return eStatus; } uint32_t HalCm_GetFreeBindingIndex( PCM_HAL_STATE state, PCM_HAL_INDEX_PARAM indexParam, uint32_t total) { CM_SURFACE_BTI_INFO surfBTIInfo; state->cmHalInterface->GetHwSurfaceBTIInfo(&surfBTIInfo); uint32_t btIndex = surfBTIInfo.normalSurfaceStart; uint32_t unAllocated = total; while (btIndex < 256 && unAllocated > 0) { uint32_t arrayIndex = btIndex >> 5; uint32_t bitMask = (uint32_t)0x1 << (btIndex % 32); if (indexParam->btArray[arrayIndex] & bitMask) { // oops, occupied if (unAllocated != total) { // clear previous allocation uint32_t allocated = total - unAllocated; uint32_t tmpIndex = btIndex - 1; while (allocated > 0) { uint32_t arrayIndex = tmpIndex >> 5; uint32_t bitMask = 1 << (tmpIndex % 32); indexParam->btArray[arrayIndex] &= ~bitMask; allocated--; tmpIndex--; } // reset unAllocated = total; } } else { indexParam->btArray[arrayIndex] |= bitMask; unAllocated--; } btIndex++; } if (unAllocated == 0) { // found slot return btIndex - total; } // no slot return 0; } void HalCm_PreSetBindingIndex( PCM_HAL_INDEX_PARAM indexParam, uint32_t start, uint32_t end) { uint32_t btIndex; for ( btIndex = start; btIndex <= end ; btIndex++) { uint32_t arrayIndex = btIndex >> 5; uint32_t bitMask = 1 << (btIndex % 32); indexParam->btArray[arrayIndex] |= bitMask; } } //*----------------------------------------------------------------------------- //| Purpose: Setup surface State with BTIndex //| Returns: Result of the operation //*----------------------------------------------------------------------------- MOS_STATUS HalCm_Setup2DSurfaceStateWithBTIndex( PCM_HAL_STATE state, int32_t bindingTable, uint32_t surfIndex, uint32_t btIndex, bool pixelPitch) { PRENDERHAL_INTERFACE renderHal = state->renderHal; MOS_STATUS eStatus; RENDERHAL_SURFACE surface; RENDERHAL_SURFACE_STATE_PARAMS surfaceParam; PRENDERHAL_SURFACE_STATE_ENTRY surfaceEntries[MHW_MAX_SURFACE_PLANES]; int32_t nSurfaceEntries, i; uint16_t memObjCtl; uint32_t offsetSrc; PRENDERHAL_STATE_HEAP stateHeap; eStatus = MOS_STATUS_UNKNOWN; nSurfaceEntries = 0; if (surfIndex == CM_NULL_SURFACE) { return MOS_STATUS_SUCCESS; } memObjCtl = CM_DEFAULT_CACHE_TYPE; // check the surfIndex if (surfIndex >= state->cmDeviceParam.max2DSurfaceTableSize || Mos_ResourceIsNull(&state->umdSurf2DTable[surfIndex].osResource) ) { CM_ASSERTMESSAGE( "Invalid 2D surface array index '%d'", surfIndex); return MOS_STATUS_UNKNOWN; } // Check to see if surface is already assigned uint32_t nBTInTable = ( unsigned char )CM_INVALID_INDEX; if ( pixelPitch ) { nBTInTable = state->bti2DIndexTable[ surfIndex ].BTI.samplerSurfIndex; } else { nBTInTable = state->bti2DIndexTable[ surfIndex ].BTI.regularSurfIndex; } if ( btIndex == nBTInTable ) { nSurfaceEntries = state->bti2DIndexTable[ surfIndex ].nPlaneNumber; stateHeap = renderHal->pStateHeap; // Get Offset to Current Binding Table uint32_t offsetDst = ( stateHeap->iCurSshBufferIndex * stateHeap->dwSshIntanceSize ) + // Points to the Base of Current SSH Buffer Instance ( stateHeap->iBindingTableOffset ) + // Moves the pointer to Base of Array of Binding Tables ( bindingTable * stateHeap->iBindingTableSize ) + // Moves the pointer to a Particular Binding Table ( btIndex * sizeof( uint32_t ) ); // Move the pointer to correct entry uint32_t *bindingTableEntry = ( uint32_t *)( stateHeap->pSshBuffer + offsetDst ); if ( pixelPitch ) { MOS_SecureMemcpy( bindingTableEntry, sizeof( uint32_t ) * nSurfaceEntries, state->bti2DIndexTable[ surfIndex ].BTITableEntry.samplerBtiEntryPosition, sizeof( uint32_t ) * nSurfaceEntries ); } else { MOS_SecureMemcpy( bindingTableEntry, sizeof( uint32_t ) * nSurfaceEntries, state->bti2DIndexTable[ surfIndex ].BTITableEntry.regularBtiEntryPosition, sizeof( uint32_t ) * nSurfaceEntries ); } return MOS_STATUS_SUCCESS; } // Get Details of 2D surface and fill the surface CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_GetSurfaceAndRegister(state, &surface, CM_ARGUMENT_SURFACE2D, surfIndex, pixelPitch)); // Setup 2D surface MOS_ZeroMemory(&surfaceParam, sizeof(surfaceParam)); surfaceParam.Type = renderHal->SurfaceTypeDefault; surfaceParam.Boundary = RENDERHAL_SS_BOUNDARY_ORIGINAL; if (!pixelPitch) { surfaceParam.bWidthInDword_UV = true; surfaceParam.bWidthInDword_Y = true; } surfaceParam.isOutput = isRenderTarget(state, surfIndex); //Cache configurations state->cmHalInterface->HwSetSurfaceMemoryObjectControl(memObjCtl, &surfaceParam); CM_CHK_MOSSTATUS_GOTOFINISH(renderHal->pfnSetupSurfaceState( renderHal, &surface, &surfaceParam, &nSurfaceEntries, surfaceEntries, nullptr)); for (i = 0; i < nSurfaceEntries; i++) { // Bind the surface State CM_CHK_MOSSTATUS_GOTOFINISH(renderHal->pfnBindSurfaceState( renderHal, bindingTable, btIndex + i, surfaceEntries[i])); } state->bti2DIndexTable[ surfIndex ].nPlaneNumber = nSurfaceEntries; // Get Offset to Current Binding Table stateHeap = renderHal->pStateHeap; offsetSrc = ( stateHeap->iCurSshBufferIndex * stateHeap->dwSshIntanceSize ) + // Points to the Base of Current SSH Buffer Instance ( stateHeap->iBindingTableOffset ) + // Moves the pointer to Base of Array of Binding Tables ( bindingTable * stateHeap->iBindingTableSize ) + // Moves the pointer to a Particular Binding Table ( btIndex * sizeof( uint32_t ) ); // Move the pointer to correct entry if ( pixelPitch ) { state->bti2DIndexTable[ surfIndex ].BTI.samplerSurfIndex = btIndex; state->bti2DIndexTable[ surfIndex ].BTITableEntry.samplerBtiEntryPosition = stateHeap->pSshBuffer + offsetSrc; } else { state->bti2DIndexTable[ surfIndex ].BTI.regularSurfIndex = btIndex; state->bti2DIndexTable[ surfIndex ].BTITableEntry.regularBtiEntryPosition = stateHeap->pSshBuffer + offsetSrc; } eStatus = MOS_STATUS_SUCCESS; finish: return eStatus; } //*----------------------------------------------------------------------------- //| Purpose: Setup Buffer surface State with BTIndex //| Returns: Result of the operation //*----------------------------------------------------------------------------- MOS_STATUS HalCm_SetupBufferSurfaceStateWithBTIndex( PCM_HAL_STATE state, int32_t bindingTable, uint32_t surfIndex, uint32_t btIndex, bool pixelPitch) { PRENDERHAL_INTERFACE renderHal = state ? state->renderHal : nullptr; MOS_STATUS eStatus; RENDERHAL_SURFACE surface; RENDERHAL_SURFACE_STATE_PARAMS surfaceParam; PRENDERHAL_SURFACE_STATE_ENTRY surfaceEntry; uint16_t memObjCtl; uint32_t offsetSrc; PRENDERHAL_STATE_HEAP stateHeap; UNUSED(pixelPitch); eStatus = MOS_STATUS_UNKNOWN; CM_CHK_NULL_RETURN_MOSERROR(state); CM_CHK_NULL_RETURN_MOSERROR(renderHal); if (surfIndex == CM_NULL_SURFACE) { return MOS_STATUS_SUCCESS; } memObjCtl = CM_DEFAULT_CACHE_TYPE; // Check to see if surface is already assigned if ( btIndex == ( uint32_t )state->btiBufferIndexTable[ surfIndex ].BTI.regularSurfIndex ) { uint32_t nSurfaceEntries = state->btiBufferIndexTable[ surfIndex ].nPlaneNumber; stateHeap = renderHal->pStateHeap; // Get Offset to Current Binding Table uint32_t offsetDst = ( stateHeap->iCurSshBufferIndex * stateHeap->dwSshIntanceSize ) + // Points to the Base of Current SSH Buffer Instance ( stateHeap->iBindingTableOffset ) + // Moves the pointer to Base of Array of Binding Tables ( bindingTable * stateHeap->iBindingTableSize ) + // Moves the pointer to a Particular Binding Table ( btIndex * sizeof( uint32_t ) ); // Move the pointer to correct entry uint32_t *bindingTableEntry = ( uint32_t *)( stateHeap->pSshBuffer + offsetDst ); MOS_SecureMemcpy( bindingTableEntry, sizeof( uint32_t ) * nSurfaceEntries, state->btiBufferIndexTable[ surfIndex ].BTITableEntry.regularBtiEntryPosition, sizeof( uint32_t ) * nSurfaceEntries ); return MOS_STATUS_SUCCESS; } // Get Details of Buffer surface and fill the surface CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_GetSurfaceAndRegister(state, &surface, CM_ARGUMENT_SURFACEBUFFER, surfIndex, 0)); // set up buffer surface MOS_ZeroMemory(&surfaceParam, sizeof(surfaceParam)); // Set isOutput by default surfaceParam.isOutput = true; //Cache configurations default state->cmHalInterface->HwSetSurfaceMemoryObjectControl(memObjCtl, &surfaceParam); // Setup Buffer surface CM_CHK_MOSSTATUS_GOTOFINISH(renderHal->pfnSetupBufferSurfaceState( renderHal, &surface, &surfaceParam, &surfaceEntry)); //Cache configurations state->cmHalInterface->HwSetSurfaceMemoryObjectControl(memObjCtl, &surfaceParam); // Bind the surface State CM_CHK_MOSSTATUS_GOTOFINISH(renderHal->pfnBindSurfaceState( renderHal, bindingTable, btIndex, surfaceEntry)); state->btiBufferIndexTable[ surfIndex ].BTI.regularSurfIndex = btIndex; state->btiBufferIndexTable[ surfIndex ].nPlaneNumber = 1; stateHeap = renderHal->pStateHeap; offsetSrc = ( stateHeap->iCurSshBufferIndex * stateHeap->dwSshIntanceSize ) + // Points to the Base of Current SSH Buffer Instance ( stateHeap->iBindingTableOffset ) + // Moves the pointer to Base of Array of Binding Tables ( bindingTable * stateHeap->iBindingTableSize ) + // Moves the pointer to a Particular Binding Table ( btIndex * sizeof( uint32_t ) ); // Move the pointer to correct entry state->btiBufferIndexTable[ surfIndex ].BTITableEntry.regularBtiEntryPosition = stateHeap->pSshBuffer + offsetSrc; eStatus = MOS_STATUS_SUCCESS; finish: return eStatus; } MOS_STATUS HalCm_Setup2DSurfaceUPStateWithBTIndex( PCM_HAL_STATE state, int32_t bindingTable, uint32_t surfIndex, uint32_t btIndex, bool pixelPitch) { MOS_STATUS eStatus; RENDERHAL_SURFACE surface; RENDERHAL_SURFACE_STATE_PARAMS surfaceParam; PRENDERHAL_INTERFACE renderHal; PRENDERHAL_SURFACE_STATE_ENTRY surfaceEntries[MHW_MAX_SURFACE_PLANES]; int32_t nSurfaceEntries, i; uint16_t memObjCtl; uint32_t offsetSrc; PRENDERHAL_STATE_HEAP stateHeap; eStatus = MOS_STATUS_UNKNOWN; renderHal = state->renderHal; if (surfIndex == CM_NULL_SURFACE) { return MOS_STATUS_SUCCESS; } memObjCtl = CM_DEFAULT_CACHE_TYPE; // Check to see if surface is already assigned uint32_t nBTInTable = ( unsigned char )CM_INVALID_INDEX; if ( pixelPitch ) { nBTInTable = state->bti2DUPIndexTable[ surfIndex ].BTI.samplerSurfIndex; } else { nBTInTable = state->bti2DUPIndexTable[ surfIndex ].BTI.regularSurfIndex; } if ( btIndex == nBTInTable ) { uint32_t nSurfaceEntries = state->bti2DUPIndexTable[ surfIndex ].nPlaneNumber; stateHeap = renderHal->pStateHeap; // Get Offset to Current Binding Table uint32_t offsetDst = ( stateHeap->iCurSshBufferIndex * stateHeap->dwSshIntanceSize ) + // Points to the Base of Current SSH Buffer Instance ( stateHeap->iBindingTableOffset ) + // Moves the pointer to Base of Array of Binding Tables ( bindingTable * stateHeap->iBindingTableSize ) + // Moves the pointer to a Particular Binding Table ( btIndex * sizeof( uint32_t ) ); // Move the pointer to correct entry uint32_t *bindingTableEntry = ( uint32_t *)( stateHeap->pSshBuffer + offsetDst ); if ( pixelPitch ) { MOS_SecureMemcpy( bindingTableEntry, sizeof( uint32_t ) * nSurfaceEntries, state->bti2DUPIndexTable[ surfIndex ].BTITableEntry.samplerBtiEntryPosition, sizeof( uint32_t ) * nSurfaceEntries ); } else { MOS_SecureMemcpy( bindingTableEntry, sizeof( uint32_t ) * nSurfaceEntries, state->bti2DUPIndexTable[ surfIndex ].BTITableEntry.regularBtiEntryPosition, sizeof( uint32_t ) * nSurfaceEntries ); } return MOS_STATUS_SUCCESS; } // Get Details of 2DUP surface and fill the surface CM_CHK_MOSSTATUS_GOTOFINISH( HalCm_GetSurfaceAndRegister( state, &surface, CM_ARGUMENT_SURFACE2D_UP, surfIndex, pixelPitch ) ); // Setup 2D surface MOS_ZeroMemory( &surfaceParam, sizeof( surfaceParam ) ); surfaceParam.Type = renderHal->SurfaceTypeDefault; surfaceParam.Boundary = RENDERHAL_SS_BOUNDARY_ORIGINAL; if ( !pixelPitch ) { surfaceParam.bWidthInDword_UV = true; surfaceParam.bWidthInDword_Y = true; } surfaceParam.isOutput = true; //Cache configurations state->cmHalInterface->HwSetSurfaceMemoryObjectControl(memObjCtl, &surfaceParam); CM_CHK_MOSSTATUS_GOTOFINISH(renderHal->pfnSetupSurfaceState( renderHal, &surface, &surfaceParam, &nSurfaceEntries, surfaceEntries, nullptr)); for (i = 0; i < nSurfaceEntries; i++) { // Bind the surface State CM_CHK_MOSSTATUS_GOTOFINISH(renderHal->pfnBindSurfaceState( renderHal, bindingTable, btIndex + i, surfaceEntries[i])); } state->bti2DUPIndexTable[ surfIndex ].nPlaneNumber = nSurfaceEntries; stateHeap = renderHal->pStateHeap; offsetSrc = ( stateHeap->iCurSshBufferIndex * stateHeap->dwSshIntanceSize ) + // Points to the Base of Current SSH Buffer Instance ( stateHeap->iBindingTableOffset ) + // Moves the pointer to Base of Array of Binding Tables ( bindingTable * stateHeap->iBindingTableSize ) + // Moves the pointer to a Particular Binding Table ( btIndex * sizeof( uint32_t ) ); // Move the pointer to correct entry if ( pixelPitch ) { state->bti2DUPIndexTable[ surfIndex ].BTI.samplerSurfIndex = btIndex; state->bti2DUPIndexTable[ surfIndex ].BTITableEntry.samplerBtiEntryPosition = stateHeap->pSshBuffer + offsetSrc; } else { state->bti2DUPIndexTable[ surfIndex ].BTI.regularSurfIndex = btIndex; state->bti2DUPIndexTable[ surfIndex ].BTITableEntry.regularBtiEntryPosition = stateHeap->pSshBuffer + offsetSrc; } eStatus = MOS_STATUS_SUCCESS; finish: return eStatus; } MOS_STATUS HalCm_SetupSampler8x8SurfaceStateWithBTIndex( PCM_HAL_STATE state, int32_t bindingTable, uint32_t surfIndex, uint32_t btIndex, bool pixelPitch, CM_HAL_KERNEL_ARG_KIND kind, uint32_t addressControl ) { MOS_STATUS eStatus; RENDERHAL_SURFACE surface; RENDERHAL_SURFACE_STATE_PARAMS surfaceParam; PRENDERHAL_INTERFACE renderHal; PRENDERHAL_SURFACE_STATE_ENTRY surfaceEntries[ MHW_MAX_SURFACE_PLANES ]; int32_t nSurfaceEntries; uint16_t memObjCtl; int32_t i; uint32_t offsetSrc; PRENDERHAL_STATE_HEAP stateHeap; UNUSED(pixelPitch); eStatus = MOS_STATUS_UNKNOWN; renderHal = state->renderHal; if ( surfIndex == CM_NULL_SURFACE ) { eStatus = MOS_STATUS_SUCCESS; goto finish; } memObjCtl = CM_DEFAULT_CACHE_TYPE; // check to see if index is valid if ( surfIndex >= state->cmDeviceParam.max2DSurfaceTableSize || Mos_ResourceIsNull( &state->umdSurf2DTable[ surfIndex ].osResource ) ) { eStatus = MOS_STATUS_INVALID_PARAMETER; CM_ASSERTMESSAGE( "Invalid 2D surface array index '%d'", surfIndex ); goto finish; } // Get Details of Sampler8x8 surface and fill the surface CM_CHK_MOSSTATUS_GOTOFINISH( HalCm_GetSurfaceAndRegister( state, &surface, kind, surfIndex, 0 ) ); // Setup surface MOS_ZeroMemory( &surfaceParam, sizeof( surfaceParam ) ); surfaceParam.Type = renderHal->SurfaceTypeAdvanced; surfaceParam.isOutput = true; surfaceParam.bWidthInDword_Y = false; surfaceParam.bWidthInDword_UV = false; surfaceParam.Boundary = RENDERHAL_SS_BOUNDARY_ORIGINAL; surfaceParam.bVASurface = ( kind == CM_ARGUMENT_SURFACE_SAMPLER8X8_VA ) ? 1 : 0; surfaceParam.AddressControl = addressControl; state->cmHalInterface->HwSetSurfaceMemoryObjectControl(memObjCtl, &surfaceParam ); renderHal->bEnableP010SinglePass = state->cmHalInterface->IsP010SinglePassSupported(); nSurfaceEntries = 0; CM_CHK_MOSSTATUS_GOTOFINISH( renderHal->pfnSetupSurfaceState( renderHal, &surface, &surfaceParam, &nSurfaceEntries, surfaceEntries, nullptr ) ); CM_ASSERT( nSurfaceEntries == 1 ); for ( i = 0; i < nSurfaceEntries; i++ ) { // Bind the surface State CM_CHK_MOSSTATUS_GOTOFINISH( renderHal->pfnBindSurfaceState( renderHal, bindingTable, btIndex + i, surfaceEntries[ i ] ) ); } stateHeap = renderHal->pStateHeap; offsetSrc = ( stateHeap->iCurSshBufferIndex * stateHeap->dwSshIntanceSize ) + // Points to the Base of Current SSH Buffer Instance ( stateHeap->iBindingTableOffset ) + // Moves the pointer to Base of Array of Binding Tables ( bindingTable * stateHeap->iBindingTableSize ) + // Moves the pointer to a Particular Binding Table ( btIndex * sizeof( uint32_t ) ); // Move the pointer to correct entry state->bti2DIndexTable[ surfIndex ].nPlaneNumber = nSurfaceEntries; state->bti2DIndexTable[ surfIndex ].BTITableEntry.sampler8x8BtiEntryPosition = stateHeap->pSshBuffer + offsetSrc; state->bti2DIndexTable[ surfIndex ].BTI.sampler8x8SurfIndex = btIndex; eStatus = MOS_STATUS_SUCCESS; finish: renderHal->bEnableP010SinglePass = false; return eStatus; } //*----------------------------------------------------------------------------- //| Purpose: Setup 3D surface State with BTIndex //| Returns: Result of the operation //*----------------------------------------------------------------------------- MOS_STATUS HalCm_Setup3DSurfaceStateWithBTIndex( PCM_HAL_STATE state, int32_t bindingTable, uint32_t surfIndex, uint32_t btIndex) { PRENDERHAL_INTERFACE renderHal = state->renderHal; MOS_STATUS eStatus; RENDERHAL_SURFACE surface; RENDERHAL_SURFACE_STATE_PARAMS surfaceParam; PRENDERHAL_SURFACE_STATE_ENTRY surfaceEntries[MHW_MAX_SURFACE_PLANES]; int32_t nSurfaceEntries, i; uint16_t memObjCtl; uint32_t offsetSrc; PRENDERHAL_STATE_HEAP stateHeap; eStatus = MOS_STATUS_UNKNOWN; nSurfaceEntries = 0; if (surfIndex == CM_NULL_SURFACE) { return MOS_STATUS_SUCCESS; } memObjCtl = CM_DEFAULT_CACHE_TYPE; // check the surfIndex if (surfIndex >= state->cmDeviceParam.max3DSurfaceTableSize || Mos_ResourceIsNull(&state->surf3DTable[surfIndex].osResource)) { eStatus = MOS_STATUS_INVALID_PARAMETER; CM_ASSERTMESSAGE( "Invalid 3D surface array index '%d'", surfIndex); return MOS_STATUS_UNKNOWN; } // Check to see if surface is already assigned uint32_t nBTInTable = (unsigned char)CM_INVALID_INDEX; nBTInTable = state->bti3DIndexTable[surfIndex].BTI.regularSurfIndex; if (btIndex == nBTInTable) { nSurfaceEntries = state->bti3DIndexTable[surfIndex].nPlaneNumber; stateHeap = renderHal->pStateHeap; // Get Offset to Current Binding Table uint32_t offsetDst = (stateHeap->iCurSshBufferIndex * stateHeap->dwSshIntanceSize) + // Points to the Base of Current SSH Buffer Instance (stateHeap->iBindingTableOffset) + // Moves the pointer to Base of Array of Binding Tables (bindingTable * stateHeap->iBindingTableSize) + // Moves the pointer to a Particular Binding Table (btIndex * sizeof(uint32_t)); // Move the pointer to correct entry uint32_t *bindingTableEntry = (uint32_t*)(stateHeap->pSshBuffer + offsetDst); MOS_SecureMemcpy(bindingTableEntry, sizeof(uint32_t)* nSurfaceEntries, state->bti3DIndexTable[surfIndex].BTITableEntry.regularBtiEntryPosition, sizeof(uint32_t)* nSurfaceEntries); return MOS_STATUS_SUCCESS; } // Get Details of 3D surface and fill the surface CM_CHK_MOSSTATUS_GOTOFINISH(HalCm_GetSurfaceAndRegister(state, &surface, CM_ARGUMENT_SURFACE3D, surfIndex, false)); // Setup 3D surface MOS_ZeroMemory(&surfaceParam, sizeof(surfaceParam)); surfaceParam.Type = renderHal->SurfaceTypeDefault; surfaceParam.Boundary = RENDERHAL_SS_BOUNDARY_ORIGINAL; //Cache configurations state->cmHalInterface->HwSetSurfaceMemoryObjectControl(memObjCtl, &surfaceParam); //Set isOutput by default surfaceParam.isOutput = true; CM_CHK_MOSSTATUS_GOTOFINISH(renderHal->pfnSetupSurfaceState( renderHal, &surface, &surfaceParam, &nSurfaceEntries, surfaceEntries, nullptr)); for (i = 0; i < nSurfaceEntries; i++) { // Bind the surface State CM_CHK_MOSSTATUS_GOTOFINISH(renderHal->pfnBindSurfaceState( renderHal, bindingTable, btIndex + i, surfaceEntries[i])); } state->bti3DIndexTable[surfIndex].BTI.regularSurfIndex = btIndex; state->bti3DIndexTable[surfIndex].nPlaneNumber = nSurfaceEntries; // Get Offset to Current Binding Table stateHeap = renderHal->pStateHeap; offsetSrc = (stateHeap->iCurSshBufferIndex * stateHeap->dwSshIntanceSize) + // Points to the Base of Current SSH Buffer Instance (stateHeap->iBindingTableOffset) + // Moves the pointer to Base of Array of Binding Tables (bindingTable * stateHeap->iBindingTableSize) + // Moves the pointer to a Particular Binding Table (btIndex * sizeof(uint32_t)); // Move the pointer to correct entry state->bti3DIndexTable[surfIndex].BTI.regularSurfIndex = btIndex; state->bti3DIndexTable[surfIndex].BTITableEntry.regularBtiEntryPosition = stateHeap->pSshBuffer + offsetSrc; eStatus = MOS_STATUS_SUCCESS; finish: return eStatus; } //|----------------------------------------------------------------------------- //| Purpose : Tag-based Synchronization on Resource //| Input : state - Hal CM State //| surface surface //| isWrite - Write or Read //| Returns : Result of the operation //|----------------------------------------------------------------------------- MOS_STATUS HalCm_SyncOnResource( PCM_HAL_STATE state, PMOS_SURFACE surface, bool isWrite) { MOS_STATUS eStatus; PMOS_INTERFACE osInterface; eStatus = MOS_STATUS_SUCCESS; osInterface = state->osInterface; if (surface == nullptr || Mos_ResourceIsNull(&surface->OsResource)) { CM_ASSERTMESSAGE("Input resource is not valid."); eStatus = MOS_STATUS_UNKNOWN; return eStatus; } osInterface->pfnSyncOnResource( osInterface, &(surface->OsResource), state->osInterface->CurrentGpuContextOrdinal, //state->GpuContext, isWrite); // Sync Render Target with Overlay Context if (surface->bOverlay) { osInterface->pfnSyncOnOverlayResource( osInterface, &(surface->OsResource), state->osInterface->CurrentGpuContextOrdinal); } return eStatus; } //! //! \brief Send Media Walker State //! \details Send MEDIA_OBJECT_WALKER command //! \param PCM_HAL_STATE state //! [in] Pointer to CM_HAL_STATE Structure //! \param PRENDERHAL_INTERFACE renderHal //! [in] Pointer to Hardware Interface Structure //! \param PMOS_COMMAND_BUFFER cmdBuffer //! [in] Pointer to Command Buffer //! \return MOS_STATUS //! MOS_STATUS HalCm_SendMediaWalkerState( PCM_HAL_STATE state, PCM_HAL_KERNEL_PARAM kernelParam, PMOS_COMMAND_BUFFER cmdBuffer) { PRENDERHAL_INTERFACE_LEGACY renderHal; MHW_WALKER_PARAMS mediaWalkerParams; MOS_STATUS eStatus; eStatus = MOS_STATUS_SUCCESS; renderHal = state->renderHal; MOS_SecureMemcpy(&mediaWalkerParams, sizeof(MHW_WALKER_PARAMS), &kernelParam->walkerParams, sizeof(CM_HAL_WALKER_PARAMS)); if (kernelParam->kernelThreadSpaceParam.threadSpaceWidth) { //per-kernel thread space is set, need use its own dependency mask mediaWalkerParams.UseScoreboard = renderHal->VfeScoreboard.ScoreboardEnable; mediaWalkerParams.ScoreboardMask = kernelParam->kernelThreadSpaceParam.globalDependencyMask; } else { //No per-kernel thread space setting, need use per-task depedency mask mediaWalkerParams.UseScoreboard = renderHal->VfeScoreboard.ScoreboardEnable; mediaWalkerParams.ScoreboardMask = renderHal->VfeScoreboard.ScoreboardMask; } eStatus = renderHal->pMhwRenderInterface->AddMediaObjectWalkerCmd( cmdBuffer, &mediaWalkerParams); return eStatus; } //! //! \brief Send GpGpu Walker State //! \details Send GPGPU_WALKER state //! \param PCM_HAL_STATE state //! [in] Pointer to CM_HAL_STATE Structure //! \param PRENDERHAL_INTERFACE renderHal //! [in] Pointer to Hardware Interface Structure //! \param PMOS_COMMAND_BUFFER cmdBuffer //! [in] Pointer to Command Buffer //! \return MOS_STATUS //! MOS_STATUS HalCm_SendGpGpuWalkerState( PCM_HAL_STATE state, PCM_HAL_KERNEL_PARAM kernelParam, PMOS_COMMAND_BUFFER cmdBuffer) { MhwRenderInterface *mhwRender; MHW_GPGPU_WALKER_PARAMS gpGpuWalkerParams; MOS_STATUS eStatus; eStatus = MOS_STATUS_SUCCESS; mhwRender = state->renderHal->pMhwRenderInterface; gpGpuWalkerParams.InterfaceDescriptorOffset = kernelParam->gpgpuWalkerParams.interfaceDescriptorOffset; gpGpuWalkerParams.GpGpuEnable = kernelParam->gpgpuWalkerParams.gpgpuEnabled; gpGpuWalkerParams.GroupWidth = kernelParam->gpgpuWalkerParams.groupWidth; gpGpuWalkerParams.GroupHeight = kernelParam->gpgpuWalkerParams.groupHeight; gpGpuWalkerParams.GroupDepth = kernelParam->gpgpuWalkerParams.groupDepth; gpGpuWalkerParams.ThreadWidth = kernelParam->gpgpuWalkerParams.threadWidth; gpGpuWalkerParams.ThreadHeight = kernelParam->gpgpuWalkerParams.threadHeight; gpGpuWalkerParams.ThreadDepth = kernelParam->gpgpuWalkerParams.threadDepth; gpGpuWalkerParams.SLMSize = kernelParam->slmSize; eStatus = mhwRender->AddGpGpuWalkerStateCmd(cmdBuffer, &gpGpuWalkerParams); return eStatus; } //! //! \brief surface Format Convert //! \details Convert RENDERHAL_SURFACE to MHW_VEBOX_SURFACE //! \param PRENDERHAL_SURFACE renderHalSurface //! [in] Pointer to RENDERHAL_SURFACE Structure //! \param PMHW_VEBOX_SURFACE_PARAMS mhwVeboxSurface //! [in] Pointer to PMHW_VEBOX_SURFACE_PARAMS //! \return MOS_STATUS //! MOS_STATUS HalCm_Convert_RENDERHAL_SURFACE_To_MHW_VEBOX_SURFACE( PRENDERHAL_SURFACE renderHalSurface, PMHW_VEBOX_SURFACE_PARAMS mhwVeboxSurface) { PMOS_SURFACE surface; MOS_STATUS eStatus = MOS_STATUS_SUCCESS; CM_CHK_NULL_GOTOFINISH_MOSERROR(renderHalSurface); CM_CHK_NULL_GOTOFINISH_MOSERROR(mhwVeboxSurface); surface = &renderHalSurface->OsSurface; mhwVeboxSurface->Format = surface->Format; mhwVeboxSurface->dwWidth = surface->dwWidth; mhwVeboxSurface->dwHeight = surface->dwHeight; mhwVeboxSurface->dwPitch = surface->dwPitch; if (surface->dwPitch > 0) { mhwVeboxSurface->dwUYoffset = ((surface->UPlaneOffset.iSurfaceOffset - surface->YPlaneOffset.iSurfaceOffset) / surface->dwPitch) + surface->UPlaneOffset.iYOffset; } mhwVeboxSurface->TileType = surface->TileType; mhwVeboxSurface->TileModeGMM = surface->TileModeGMM; mhwVeboxSurface->bGMMTileEnabled = surface->bGMMTileEnabled; mhwVeboxSurface->rcMaxSrc = renderHalSurface->rcMaxSrc; mhwVeboxSurface->pOsResource = &surface->OsResource; finish: return eStatus; } //! //! \brief Set Vtune Profiling Flag //! \details Trun Vtune Profiling Flag On or off //! \param PCM_HAL_STATE state //! [in] Pointer to CM_HAL_STATE Structure //! \return MOS_STATUS_SUCCESS //! MOS_STATUS HalCm_SetVtuneProfilingFlag( PCM_HAL_STATE state, bool vtuneOn) { state->vtuneProfilerOn = vtuneOn; return MOS_STATUS_SUCCESS; } //*----------------------------------------------------------------------------- //| Purpose: Get the offset for the Task Sync Location given the task ID //| Returns: Sync Location //*----------------------------------------------------------------------------- int32_t HalCm_GetTaskSyncLocation( PCM_HAL_STATE state, int32_t taskId) // [in] Task ID { return (taskId * state->cmHalInterface->GetTimeStampResourceSize()); } void HalCm_GetLegacyRenderHalL3Setting( CmHalL3Settings *l3SettingsPtr, RENDERHAL_L3_CACHE_SETTINGS *l3SettingsLegacyPtr ) { *l3SettingsLegacyPtr = {}; l3SettingsLegacyPtr->bOverride = l3SettingsPtr->overrideSettings; l3SettingsLegacyPtr->bEnableSLM = l3SettingsPtr->enableSlm; l3SettingsLegacyPtr->bL3CachingEnabled = l3SettingsPtr->l3CachingEnabled; l3SettingsLegacyPtr->bCntlRegOverride = l3SettingsPtr->cntlRegOverride; l3SettingsLegacyPtr->bCntlReg2Override = l3SettingsPtr->cntlReg2Override; l3SettingsLegacyPtr->bCntlReg3Override = l3SettingsPtr->cntlReg3Override; l3SettingsLegacyPtr->bSqcReg1Override = l3SettingsPtr->sqcReg1Override; l3SettingsLegacyPtr->bSqcReg4Override = l3SettingsPtr->sqcReg4Override; l3SettingsLegacyPtr->bLra1RegOverride = l3SettingsPtr->lra1RegOverride; l3SettingsLegacyPtr->dwCntlReg = l3SettingsPtr->cntlReg; l3SettingsLegacyPtr->dwCntlReg2 = l3SettingsPtr->cntlReg2; l3SettingsLegacyPtr->dwCntlReg3 = l3SettingsPtr->cntlReg3; l3SettingsLegacyPtr->dwSqcReg1 = l3SettingsPtr->sqcReg1; l3SettingsLegacyPtr->dwSqcReg4 = l3SettingsPtr->sqcReg4; l3SettingsLegacyPtr->dwLra1Reg = l3SettingsPtr->lra1Reg; return; } uint64_t HalCm_ConvertTicksToNanoSeconds( PCM_HAL_STATE state, uint64_t ticks) { if (state->tsFrequency == 0) { // if KMD doesn't report an valid value, fall back to default configs return state->cmHalInterface->ConverTicksToNanoSecondsDefault(ticks); } return (ticks * 1000000000) / (state->tsFrequency); } //! //! \brief Check GPU context //! \details Check if the GPU context is valid for CM layer //! \param MOS_GPU_CONTEXT gpuContext //! [in] GPU Context ordinal //! \return true/false //! bool HalCm_IsValidGpuContext( MOS_GPU_CONTEXT gpuContext) { if( gpuContext == MOS_GPU_CONTEXT_RENDER3 || gpuContext == MOS_GPU_CONTEXT_RENDER4 || gpuContext == MOS_GPU_CONTEXT_CM_COMPUTE || gpuContext == MOS_GPU_CONTEXT_VEBOX) { return true; } else { CM_ASSERTMESSAGE("Invalid GPU context for CM."); return false; } }